https://docs.python.org/3/c-api/init.html#c.Py_EndInterpreter states that "Py_FinalizeEx() will destroy all sub-interpreters that haven’t been explicitly destroyed at that point."

As discussed in https://github.com/hexchat/hexchat/issues/2237, Python 3.7+ doesn't currently do that - it calls Py_FatalError instead.

That change came from https://github.com/python/cpython/pull/1728, which was based on my initial PEP 432 refactoring work, and I didn't realise that implicitly cleaning up lingering subinterpreters was a documented behaviour.

So I think we should just fix it to behave as documented, and add a new regression test to make sure it doesn't get broken again in the future.