Mutex stays locked when a stop races with unlock #66

Mutex::LockAwaiter can throw CancelledException after unlock() has already handed it the lock, leaving the mutex locked forever.

1. A coroutine is suspended in co_await mutex.lock().
2. Its stop is requested. The cancellation is posted and has not run yet.
3. The holder calls unlock(). resume_now takes the waiter and resumes it; locked_ stays true, as ownership is passed to the waiter.
4. await_resume sees token.stop_requested() and throws.

Nobody owns the lock now, so locked_ is never cleared and every later lock() waits forever.

This predates #65 (the old check token.stop_requested() && !(active && *active) behaves the same). A fix could be to throw only when the cancellation resumed the waiter (e.g. a cancelled flag set in the register_cancel action, as the other awaiters do), or to unlock before throwing.