The pool herder didn't signal the thread when decimating the flock,
causing the thread to be leaked.

Spotted by: xcir
Fixes: #1490

Also close a couple of other races:

When pthread_cond_timedwait returns ETIMEDOUT, the predicate
(wrk->task.func == NULL) could still have changed while reacquiring
the mutex. If so, the signal would've been lost and the thread
leaked. Changed the idle wait loop in Pool_Work_Thread() to always
check the predicate before resuming the cond_wait.

The update of the predicate (wrk->task.func) from e.g. Pool_Task() is
done without holding the mutex. This opens a race on the predicate,
and the possibility of the worker going back on cond_wait loosing the
signal. Changed to update the predicate while holding the mutex.