The old code collected a list of THD's, locked the THD's from getting
deleted by locking two mutex and then later in a separate loop
sent a kill signal to each THD.

The problem with this approach is that, as THD's can be reused,
the second time the THD is killed, the mutex can be taken in
different order, which signals failures in safe_mutex.

Fixed by sending the kill signal directly and not collect the THD's
in a list to be signaled later.  This is the same approach we are using
in kill_zombie_dump_threads().

Other things:
- Reset safe_mutex_t->locked_mutex when freed (Safety fix)