This fixes threaded coredumps and streamlines the code.  The old code
caused crashes and hung coredumps.  The new code has been tested for
some time already and appears to be robust.  Changes:

 - the code now uses completions instead of a semaphore and a waitqueue,
   attached to mm_struct:

        /* coredumping support */
        int core_waiters;
        struct completion *core_startup_done, core_done;

 - extended the completion concept with a 'complete all' call - all pending
   threads are woken up in that case.

 - core_waiters is a plain integer now - it's always accessed from under
   the mmap_sem. It's also used as the fastpath-check in the sys_exit()
   path, instead of ->dumpable (which was incorrect).

 - got rid of the ->core_waiter task flag - it's not needed anymore.