GeezerVison

(This should solve the semi-frequent buildbot failures we've been seeing.)

In Py_EndInterpreter() we almost immediately wait for all non-daemon Python threads to finish, by calling theading._shutdown(). However, with a per-interpreter GIL there's a race because threading._shutdown() doesn't wait long enough. This change fixes that by making sure _PyThreadState_DeleteCurrent() completely finishes before releasing the lock on which threading._shutdown() depends.

The gist of it is that we release that lock (and clean it up) later than we would normally be able to, by doing the cleanup in another thread via a "pending" call.

This isn't an ideal solution, but other approaches I tried involved more invasive changes, which I'd like to avoid this close to the beta 1 release.

• Issue: test__xxsubinterpreters is Occasionally Crashing #104341