bug: Windows CI flake - pytest-xdist workers race on the shared torchinductor cache

The windows job fails intermittently on main and on PRs with a PermissionError on a file under torchinductor_runneradmin, in whichever torch-based test happens to lose the race. It is not tied to a particular test or to the change under review.

Occurrences:

• PR #9054, test_torch_index_with_nans[Legacy], worker gw0 - https://github.com/lance-format/lance/actions/runs/34195246750/job/101961424216
• main, test_torch_index_with_nans[V3], worker gw3 - https://github.com/lance-format/lance/actions/runs/33933386340/job/101216448469
• main, test_index_cast_centroids, worker gw4 - https://github.com/lance-format/lance/actions/runs/33893404408/job/101090259544

torch._inductor.exc.InductorError: PermissionError: [Errno 13] Permission denied:
'C:\Users\RUNNER~1\AppData\Local\Temp\torchinductor_runneradmin\5h\c5h7jepkg4o4vxj5qik5e3dax3tsrhzgga3lujk3z74ctjdxcg3i.py'

The first two hit the identical cache path. That filename is a content hash of the generated kernel, so both runs were compiling the same kernel, and every failure so far names a different worker. torchinductor_runneradmin is a single per-user directory shared by all pytest-xdist workers, and on Windows a file held open by one process cannot be replaced by another, so two workers compiling the same kernel collide instead of harmlessly rewriting it.

A per-worker cache directory should remove the collision, e.g. deriving TORCHINDUCTOR_CACHE_DIR from PYTEST_XDIST_WORKER in the Windows test step.

Related: #8789 listed a different Windows temp race (test_build_ivf_flat::case_3) and is closed.

Start with the Windows test step in the GitHub Actions configuration and review the listed failing runs to confirm the shared torchinductor cache behavior. Configure a cache directory derived from PYTEST_XDIST_WORKER, then rerun the Windows pytest-xdist job and verify that concurrent workers no longer collide.