gh-158090: Fix hang at shutdown when a daemon thread resets the GIL drop request

[lazerg]

When a daemon thread in take_gil() sees that it must exit, it released gil->mutex before clearing the drop request it had set on the GIL holder. In that gap the holder could run drop_gil(), still see the bit set, and wait on switch_cond forever, since the only waiter was about to hang. Clearing the bit before releasing the mutex closes the window: the holder needs gil->mutex to drop the GIL, so it now always sees the bit already cleared.

I checked it with the lldb recipe from the issue on a --with-pydebug build. Before the change the process hangs with the main thread in drop_gil(). After it, stopping the daemon thread right after the unlock no longer hangs, and the process exits normally.

• Issue: Hang at shutdown: take_gil() clears the GIL drop request after releasing gil->mutex #158090