A new patch:
- I replaced "_active_limbo_lock.release()" by
"_active_limbo_lock=_allocate_lock()"

- I replaced the successive deletions in _active by a re-creation with
only the current thread. There is no difference in the result, but I
agree that the intent is more clear.

- yes, the main thread is marked as done when the interpreter exits
(hence the convoluted tests with subprocesses): in Modules/main.c,
WaitForThreadShutdown() calls threading._shutdown().

Also, I hope the tests make more sense now.