intermittent ResourceTracker called reentrantly for resource cleanup in tests

[graingert]

Bug report

Bug description:

I occasionally see

UserWarning: ResourceTracker called reentrantly for resource cleanup, which is unsupported. The semaphore object '/mp-gtkzf9rd' might leak.

warned in the test logs, I suspect test_force_shutdown_workers_dead_workers is the culprit.

It looks like the ResourceTracker was built to warn on re-entrant execution but I think it can be made to work instead.

CPython versions tested on:

CPython main branch

Operating systems tested on:

No response

Linked PRs

• gh-131788: make resource_tracker re-entrant safe #131787
• [3.14] gh-131788: make resource_tracker re-entrant safe (GH-131787) #137737
• [3.13] gh-131788: make resource_tracker re-entrant safe (GH-131787) #137738

[gpshead]

heads up that backports seem to have happened and the PR shipped in 3.13.7 and at least one person says it is causing unspecified as of yet issues cc: @ngoldbaum - #137738 (comment)

my own comment on the PR many moons ago when approving did question if this was safe enough to backport, but the automation was still in place so those understandably happened regardless (it has not landed in 3.14 yet due to rc phase). #131787 (comment)

[ngoldbaum]

It looks like the backport of the fix for this issue is causing problems in the loky project, which is used by joblib (in turn used by lots of stuff, including scikit-learn).

joblib/loky#459

joblib/joblib#1738

[geofft]

I think we have another couple of reports of this regression over at astral-sh/python-build-standalone#774. Specifically note that all the reports are that this is failing on Windows with an ImportError for _posixsubprocess, which is imported by multiprocessing.util.spawnv_passfds(), the call to which moved around in this PR.

Did something on this PR cause code to run on Windows that was previously getting skipped (or ImportError to be swallowed)?