test(hip-kernel-provider): port mlops engine integration tests

[adickin-amd]

Summary

Ports hip-kernel-provider's mlops-engine-local Batchnorm/RMSnorm/Layernorm integration tests into the shared dnn-providers/integration-tests suite and wires the provider onto add_external_integration_test_target() plus the YAML-driven quick/standard/comprehensive/full CTest tier framework, mirroring the pattern already merged for miopen-provider. JIRA ID : ALMIOPEN-2283

This covers the mlops-engine half of ALMIOPEN-2283 only; the story also asks for the ASM SDPA engine and rocKE is explicitly deferred there, but this PR intentionally leaves both untouched (rocKE per the story, ASM SDPA per direction from the assignee) to keep the change scoped to one engine.

Running the full (unfiltered) common suite against HIP_MLOPS_ENGINE for the first time - exactly the "does the external test cmake stuff all work as expected" unknown the story calls out - surfaced one real, pre-existing bug unrelated to mlops itself: IntegrationGpuConvFwdSerializeRoundTripFp32 bypasses the harness's usual engine-support check and asserts graphObj.build() succeeds unconditionally, so it fails outright (instead of skipping) against any engine that doesn't implement conv. Fixed by extracting the existing check into a reusable checkEngineSupportOrSkip() harness helper and calling it from that test too.

Risk Assessment

Medium risk. The diff itself is test/build-only - no production or kernel code changed - and is thoroughly covered locally (full unfiltered suite: 3939 tests, 0 failures on gfx1151). However, it activates previously-inactive generic integration coverage (Batchnorm, RMSNorm) against HIP_MLOPS_ENGINE through a new entrypoint and adds net-new generic coverage (RMSNorm backward, Layernorm forward), so per policy it needs a full multi-arch sweep before merge - this exact engine/op family has a recent history of arch-specific failures (#9139, #9151, both already fixed upstream and confirmed passing on this branch) that a sweep could plausibly resurface on unswept targets, and running the full suite here already surfaced one real (now-fixed) common-suite gap.

ASIC Coverage

Full multi-arch sweep required. This activates generic integration test cases (Batchnorm, RMSNorm, Layernorm) in a new provider lane (HIP_MLOPS_ENGINE via hip-kernel-provider-external-integration-check) and adds two net-new generic test files; a newly activated or added generic case can fail on any target, so standard PR CI (routine families: gfx94X, gfx950 on Linux, gfx1151 on Windows) is not sufficient on its own. The full supported family set (gfx94X-all, gfx950-all, gfx110X-all, gfx120X-all, plus gfx1151) should pass before merge.

Testing Summary

• Local superbuild of the hip-kernel-provider preset builds clean under -Werror (699/699 targets).
• Ran the entire hipdnn_integration_tests suite unfiltered, pinned to HIP_MLOPS_ENGINE via --test-engine, on local gfx1151 hardware: 3939 tests, 1692 passed, 2247 skipped (ops HIP_MLOPS_ENGINE doesn't implement + golden-ref CPU-only cases + the intentional inference-mode mixed-precision layernorm skip), 0 failed.
• That full run surfaced and led to fixing a pre-existing common-suite bug (conv serialize round-trip skipping engine-support checks - see Technical Changes); re-ran after the fix with identical pass count and 0 failures, confirming no regression.
• Confirmed two previously-filed known failures for this exact engine/op combination (hip-kernel-provider mlops engine backnorm bwd integration test fails #9139 batchnorm backward NHWC/NDHWC, hipdnn integration tests fail due to hip-kernel-provider CalcStats remainder loop #9151 CalcStats remainder loop) already pass on this branch - pre-existing bugs fixed upstream, not caused by or fixed by this PR.
• Reconfirmed identical pass/skip/fail counts after the pre-commit clang-format pass and after the yaml-file rename.

Testing Checklist

• Local superbuild - cmake --build build --target hipdnn_integration_tests hip_kernel_provider_integration_tests (hip-kernel-provider preset, -Werror) - Status: Passed
• hip-kernel-provider external integration tests (HIP_MLOPS_ENGINE), full unfiltered suite - hipdnn_integration_tests --test-engine HIP_MLOPS_ENGINE --test-config config/HIP_MLOPS_ENGINE.toml - ASICs: gfx1151 - Status: Passed (3939 ran, 1692 passed, 2247 skipped, 0 failed)
• Multi-arch sweep - TheRock multi-arch CI - ASICs: gfx94X, gfx950, gfx110X, gfx120X - Status: Pending
• PR CI - GitHub PR checks - Status: Pending

Technical Changes

• dnn-providers/integration-tests: Batchnorm and RMSNorm-forward already existed generically and are a strict superset of the mlops-local coverage (1D support, full/regression tiers, golden references, parametrized backward-activation modes); no new code needed for those.
• dnn-providers/integration-tests: adds integration_tests/rmsnorm/IntegrationGpuRMSNormBackward.cpp, reusing the existing RMSNormTestCase/getRMSNormTestCases() generator (bias/isTraining are forward-only concerns that rmsnorm_backward() ignores).
• dnn-providers/integration-tests: adds common/LayernormCommon.hpp and integration_tests/layernorm/IntegrationGpuLayernormForward.cpp - net new, no layernorm coverage existed anywhere in the common suite before this.
• dnn-providers/integration-tests: extends the harness's tolerance dispatch (IntegrationGraphVerificationHarness.hpp) with RMSNormBackwardNode/LayerNormNode entries, both reusing existing hipdnn_test_sdk tolerance functions.
• dnn-providers/integration-tests: extracts the harness's existing engine-support-check block out of verifyGraph() into a reusable checkEngineSupportOrSkip() helper (no behavior change for existing callers) and calls it from IntegrationGpuConvForwardSerializeRoundTrip.cpp, which was bypassing that check entirely and failing (rather than skipping) against engines that don't implement conv, such as HIP_MLOPS_ENGINE.
• dnn-providers/hip-kernel-provider: removes the now-ported src/integration_tests/hip_mlops_engine/ tree (Batchnorm/, RMSnorm/, Layernorm/, local ActivationCommon.hpp) and its CMake wiring. The provider-local IntegrationGraphVerificationHarness.hpp and asm_sdpa_engine/ tests are untouched, since that harness is still shared with the (out of scope) ASM SDPA engine.
• dnn-providers/hip-kernel-provider: adds config/HIP_MLOPS_ENGINE.toml (template only, no known skips or tolerance overrides yet) and wires add_external_integration_test_target() (ENGINE_NAME HIP_MLOPS_ENGINE, PLUGIN_TARGET hip_kernel_provider).
• dnn-providers/hip-kernel-provider: adds HIP_MLOPS_ENGINE_test_categories.yaml, HIP_MLOPS_ENGINE_test_categories_external.yaml, and HIP_MLOPS_ENGINE_test_categories_integration.yaml, adopting miopen-provider's tiering. Files are engine-name-prefixed (rather than the generic test_categories*.yaml miopen/hipblaslt use) since hip-kernel-provider has more than one engine (ASM_SDPA_ENGINE, HIP_MLOPS_ENGINE) and a future per-engine external-integration-check target shouldn't collide with these names.

[adickin-amd]

Note: not ready to review, just putting up to see if it builds on CI and what needs to change on the yaml filters.

[therock-pr-bot]

❌ PR Check — Action Required

Check	Status	Details
🌿 Branch Name	✅ Pass	—
📝 PR Title/Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	❌ Fail	Error: Source/code files changed without an accompanying unit test. Expected: add at least one test file named like test_<name>.py / test_<name>.cpp (or <name>_test.*). Current: code file(s) changed: dnn-providers/integration-tests/src/common/LayernormCommon.hpp, dnn-providers/integration-tests/src/harness/IntegrationGraphVerificationHarness.hpp, dnn-providers/integration-tests/src/integration_tests/conv/IntegrationGpuConvForwardSerializeRoundTrip.cpp, dnn-providers/integration-tests/src/integration_tests/layernorm/IntegrationGpuLayernormForward.cpp, dnn-providers/integration-tests/src/integration_tests/rmsnorm/IntegrationGpuRMSNormBackward.cpp; no test file found
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

⚠️ 1 policy check(s) failed. Please address the issues above before this PR can be Reviewed.

🚫 Please fix the failed policies

• ❌ Unit Test

The Not ready to Review label was added to this PR. Once all policies pass, the label is removed automatically.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

[therock-pr-bot]

🚫 Please fix the failed policies before requesting reviews.

The following policy checks failed:

• ❌ Unit Test

The Not ready to Review label has been added to this PR.
Once all policies pass, the label will be removed automatically.