[MXFP8] Fix compile test configuration - #5809

Summary (author-provided intent, AI-edited)

Improve MXFP8 compile-test execution on SM103 by selecting the supported
weight-gradient accumulation mode. This test-only change also applies to
other SM10x devices and does not add hardware support.

Why

AI-assisted technical notes: test_mxfp8_linear_runs_outside_plain_eager
compiles an MXFP8Linear with its default inplace_wgrad_accum=True.
Traced backward explicitly rejects that configuration, so the two compiled
cases fail before checking the weight gradient. The compile cases need the
supported non-in-place path; the checkpoint cases can retain their current path.

What changed

AI-assisted technical notes: Set linear.inplace_wgrad_accum = False only
when this test chooses torch.compile. This changes the test setup and does
not change the production configuration or the checkpoint cases.

Testing

AI-assisted test notes: On B300 (SM 10.3), with PyTorch
2.16.0.dev20261102+cu130 and Triton 3.9.0, the two compile cases on base
485c065cb248863bf2e2c65588becd6d7a8bcf83 fail at the in-place WGRAD guard.
With this change, the four compile/checkpoint combinations pass, and the
complete test file reports 25 passed. The tested file was a byte-identical
copy of the changed file.

python -m pytest -p no:cacheprovider -q candidate_test_mxfp8_linear.py::test_mxfp8_linear_runs_outside_plain_eager
# 4 passed
python -m pytest -p no:cacheprovider -q candidate_test_mxfp8_linear.py
# 25 passed

The local pre-commit run --all-files run passed the formatting, lint and
other available hooks; the link checker skipped because lychee was absent.
Pyrefly could not pass locally because that environment lacked the project
dependencies. The same Pyrefly 0.45.1 check with project dependencies on an
isolated source copy completed with 0 errors and 2 warnings.