Add ONNX detector runtime pipeline

[milliyin]

Summary

Add a reference ONNX Runtime deployment path for TR-Hash Vision v8 detector exports.

This adds a small complexity.deploy.onnx_detector package plus a CLI example that runs the exported raw ONNX tensors through the full deployment contract:

• image preprocessing and letterbox restoration
• ONNX Runtime inference
• shared DFL box decoding
• confidence filtering for both branches
• class-aware NMS for O2M only
• NMS-free top-score filtering without NMS
• stable JSON detection schema with normalized boxes, pixel boxes, class IDs, and scores
• provider selection/reporting for CPU, CUDA, and TensorRT

Docs

Adds:

• docs/onnx_deploy_contract.md
• docs/onnx_deploy.md

The docs include the raw tensor contract, grid/DFL decode rules, branch behavior, and runnable CPU/CUDA CLI examples.

Testing

Passed:

python -m pytest -q tests\test_onnx_detector_skeleton.py tests\test_onnx_detector_core.py tests\test_onnx_detector_metadata.py tests\test_onnx_detector_pipeline.py tests\test_onnx_detect_cli.py
16 passed

Issue closed #13