Bug: TracerWarning leaks to stderr on winml export, build/hf.py paths

[tezheng]

Summary

After PR #398 removes the eager TracerWarning filter from `_warnings.py` (the root-cause fix for the CLI startup regression), three call sites that invoke `torch.onnx.export` (transitively via `export_onnx`/`export_pytorch`) now emit `TracerWarning` to stderr because they have no local `warnings.catch_warnings()` wrapper:

Path	Wrapped?
`commands/build.py:1060-1061`	✅ Yes (the wrapping that #398 relies on)
`commands/export.py:347` (`winml export` direct CLI)	❌ No
`build/hf.py:216` (called by `winml.modelkit.build.hf.build_hf_model()`)	❌ No
`commands/perf.py` (perf invocation paths)	❌ No

Why not in #398

#398's core change is reverting the eager `from torch.jit import TracerWarning` because it imported all of torch (~1.7s) at startup. Wrapping each call site is a separate concern — and the cleanest fix is "wrap once inside `export_pytorch()`" rather than touching every caller.

Recommended fix

Move the `with warnings.catch_warnings(): warnings.filterwarnings("ignore"):` block from `commands/build.py:1060-1061` into `src/winml/modelkit/export/pytorch.py::export_pytorch()` so all callers get the same TracerWarning suppression by default. Removes:

• The duplication (no need to re-wrap at every caller)
• The "remember to add catch_warnings" footgun for new callers
• The "_warnings.py says build.py wraps it" claim that's only partially true

Severity

Low — UX/stderr noise. Functional behavior unaffected. Acknowledged as a known caveat in #398.

Reported by @DingmaomaoBJTU and @timenick in code review of #398.

[DingmaomaoBJTU]

Triage summary

• Component: Load & Export
• Priority: P2
• Owner(s): @vortex-captain @chinazhangchao

TracerWarning leaks to stderr from unwrapped export call sites (export.py, build/hf.py, perf.py); move the warnings.catch_warnings() block into export_pytorch() so all callers are covered.