Can not communicate with subprocess during interpreter shutdown on windows

[tacaswell]

Bug report

Bug description:

In #104826 (backported to 3.12 in #105277) to fix #104690 disallowed creating new threads during interpreter finalization. This poses a problem on windows because the Popen._communicate uses a thread (see matplotlib/matplotlib#27437 (comment) for discussion).

In Matplotlib we are holding onto a latex process and triggering this via a weakref finalize, but a minimal reproducer is:

from subprocess import Popen, PIPE
import atexit

# anything that holds stdin and stdout should work
proc = Popen(['powershell'], stdin=PIPE, stdout=PIPE)

def cleanup():
    print('hi bob')
    proc.kill()
    proc.communicate()

atexit.register(cleanup)

xref matplotlib/matplotlib#27437

CPython versions tested on:

3.12

Operating systems tested on:

Windows

[serhiy-storchaka]

cc @chgnrdv, @gpshead

[zooba]

(cc @ericsnowcurrently)

One possibility is to transform the Windows code here into a purely native thread, rather than a Python thread, but that is likely to break people who are passing file objects to stdout=/stderr= besides our constants. That's probably unlikely, but we also shouldn't break it.

What might be better, and why I pinged Eric, is to find a finalization ordering where we can run atexit callbacks and then require all threads to be exited. That may not help with the similar GC issues, but it seems like atexit ought to be okay with running a bit earlier.

[ericsnowcurrently]

See gh-116677.