CI: pytest-xdist in every suite, real-mpirun tests rebalanced to the kvcache job; version 4.0.1

[FileSystemGuy]

CI wall-clock has crept up to about 3:45 per run. The critical path was the others job: its test step alone took 174s (main suite + package suite + vdb, serial), while the kvcache job finished a minute earlier. Two things were true that the layout did not exploit: four real-mpirun tests are 35% of the main suite when run serially (a fixed ~10s quiesce window each), and both heavy suites parallelize cleanly.

Changes

• pytest-xdist added to the test extra (pyproject.toml, uv.lock). Every suite with more than a few seconds of work runs with -n auto (one worker per runner vCPU). No test code changed; both heavy suites were already xdist-clean.
• Real-mpirun probe file moved next to kvcache. kvcache runs kv_cache_benchmark/tests then tests/integration/test_shared_fs_probe_real_mpi.py; others runs tests with that file --ignored, then mlpstorage_py/tests, then vdb_benchmark/tests. Every test still runs exactly once per workflow run; the comment in the workflow says so.
• Version bumped 3.0.46 → 4.0.1 (pyproject.toml, uv.lock) — first release on the v4.0 line. RULES_EDITION stays 3.0 (tool releases never bump the edition; see Rules.md "Rules editions and comparability classes").

Measurements (this box, 4 workers = the runner's vCPU count)

Suite	Serial	-n 4
tests minus real-MPI	~75s	29s
real-MPI file (4 tests)	42s	11s
kv_cache_benchmark/tests	84–112s	28–38s
mlpstorage_py/tests	19s	18s (no gain; harmless)
vdb_benchmark/tests	1s	serial, unchanged

Stability: kvcache and the real-MPI file each ran 3× under 4 workers, clean every time. One kvcache test (TestThreeTierEvictionCascade::test_demote_path_preserves_data) failed once earlier while this machine was running four serial pytest processes at once, and passed on every xdist run; noting it here as a test to watch if the kvcache job ever goes red.

Expected CI: both jobs' test steps land around 40–60s, so the run should be bounded by setup (checkout, apt OpenMPI, uv sync) plus about a minute, roughly 2:00 instead of 3:45.

Not in this PR

• Caching (setup-uv enable-cache: true, apt) would trim the ~45–60s setup further.
• The workflow comment about "2-vcpu runners" was stale (public-repo ubuntu-latest has 4); rewritten.

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅