Add HybridModel Python recipe API and launcher support

[Phlip79]

Summary

• Add a declarative Python recipe API for HybridModel, including model-wide defaults, per-layer configuration, embedding/loss markers, validation, and lowering.
• Support loading recipes from dotted modules or Python files and constructing models directly with HybridModel(config=recipe).
• Add --model-recipe launcher and ModelBuilder integration while preserving the legacy hybrid string DSL.
• Add a Nemotron-3 Nano recipe, launcher examples, unit/equivalence coverage, and an end-to-end functional workload.

This consolidates the changes from #4537, #5030, #5031, #4538, and #4539 into one PR.

Upstream sync

The seven substantive commits were replayed onto main at a54bd8cc1; eight stack-only merge commits were omitted. Conflict resolutions preserve current-main behavior for:

• explicit ProcessGroupCollection and logging process-group plumbing
• legacy HybridModel virtual pipeline parallelism, MLA down-projection fusion, and CUDA graph handling
• GTP-aware total model-size calculation
• strict functional-test shell handling and the existing quarantined Nemotron Pico workload

All commits are DCO-signed-off and GitHub-verified.

Validation

• git diff --check upstream/main...HEAD
• Check-only formatter: Black 26.3.0, isort, pylint (10.00/10), and Ruff
• HybridModel test collection: 144 tests collected
• 8-GPU HybridModel unit suite (this host has no usable NVIDIA driver)
• H100 functional workload and golden generation; golden_values_dev_dgx_h100.json is not yet present

The cached lint image did not include mypy; the repository formatter treats that invocation as optional.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.