bug: WinMLAutoModel.from_onnx uses incorrect cache dir/file name

[xieofxie]

After #659, verified by --no-skip-build

• WinMLAutoModel uses output_dir = get_model_dir(onnx_path.stem, cache_dir=cache_dir_path) instead of hash of model file / path
• build_onnx_model uses stem instead of hash of config winml_build_config

[DingmaomaoBJTU]

Triage summary

• Component: Build
• Priority: P1

WinMLAutoModel.from_onnx and build_onnx_model use onnx_path.stem instead of a hash of the model file/path and build config for cache directory naming. This can cause cache collisions when different models share the same filename stem, or when the same model is built with different configs. Suggested fix: use a content/path hash for cache key derivation.

[DingmaomaoBJTU]

Triage Summary

Priority: P2
Component: core (build/cache)
Owner: @chinazhangchao

Analysis: WinMLAutoModel.from_onnx uses onnx_path.stem for cache directory naming instead of a hash-based key. This means different ONNX models with the same filename will collide in cache, and config changes won't invalidate the cache. Introduced after PR #659.

Action: Replace stem-based cache dir naming with content-hash or config-hash to ensure cache correctness. Verify with --no-skip-build flag.