feat: add a tract backend for classical PaddleOCR with a bounded shape-keyed plan cache

[Haoxincode]

What is the proposed feature?

Complete the tract follow-up left by #1275 and #1282 for classical PaddleOCR:

• Route DBNet, CRNN, and AngleNet in xberg-paddle-ocr through the shared InferenceBackend seam.
• Add a bounded shape-keyed tract plan cache for DBNet and CRNN dynamic inputs.
• Expose an explicit no-ORT Cargo feature, for example paddle-ocr-tract (exact naming is up to the maintainers).
• Keep the existing paddle-ocr / ONNX Runtime path as the native default.
• Add ORT/tract parity tests for PP-OCRv5 and PP-OCRv6 fixtures.
• Document cache limits, eviction behavior, memory use, warm latency, and unsupported model combinations.

Suggested acceptance criteria:

• The no-ORT PaddleOCR feature's dependency tree does not contain ort.
• PP-OCRv5/v6 detection, classification, and recognition pass fixed corpus tests.
• Text, boxes, confidence, and reading-order differences from the ORT baseline have documented tolerances.
• The plan cache has a capacity or memory budget, deterministic eviction behavior, and documented concurrency semantics.
• Shape misses, unsupported models, and resource-limit failures return explicit errors or diagnostics.
• Benchmarks cover first load, cache hit, cache miss, peak memory, and CPU latency.

Why would this be a good addition?

The v1.0.5 pure-Rust inference documentation says DBNet and CRNN already achieve strong numerical parity with ONNX Runtime, while the remaining blocker is dynamic-shape plan reuse. AngleNet is fixed-shape and can already run through tract.

The current xberg-paddle-ocr crate still depends directly on ort and stores ort::session::Session, so users cannot deploy classical PP-OCR as a pure-Rust Rust crate even when GPU execution providers are not needed.

A separate explicit feature would support simpler offline and embedded Rust distribution without changing the existing native default. This request is intentionally limited to classical PaddleOCR; TATR, PP-DocLayout-V3, and SLANeXT have separate tract operator/model constraints.

Related context:

• feat: use tract as a pure-Rust ONNX backend to unlock WASM/Android for ORT-gated models #1275 originally included routing xberg-paddle-ocr through the inference seam.
• feat(inference): pure-Rust tract ONNX backend for WASM, Android, iOS (#1275, #1276) #1282 delivered the common tract backend and explicitly recorded PaddleOCR's shape-keyed plan cache as follow-up work.
• Current documentation: https://docs.xberg.io/concepts/tract-inference/

[Goldziher]

Implemented and merged. 2d335a1449 runs classical PaddleOCR on the tract backend; 087396d03f pins the tract DBNet plans per page extent rather than a fixed canvas.

Worth recording, because the fixed-canvas approach in the original plan is the wrong one: padding to a fixed canvas breaks DBNet, since the SE blocks' GlobalAveragePool reduces over the whole padded extent — 827 of 307200 pixels flipped and 29 words were lost. The shape-keyed plan cache described in this issue was correct; tract needs concrete, not square, shapes. Final parity is 5/0/0 with detection IoU 1.000 and 430=430 words.

Documentation in CHANGELOG follows with the v1.1.0 release.