Fix extremely rare case of infinite send lost signals if thread restarted by SIGQUIT after restart signal is lost

[ivmai]

Source: master (02c9d85)

This is follow up of #630, #633.
As mentioned in the code

bdwgc/pthread_stop_world.c

Line 1257 in d07d2ff

/* FIXME: Still, an extremely low chance exists that the */

  STATIC int GC_restart_all(void)
  {
    int n_live_threads = 0;
    ...
          if (GC_retry_signals
                && AO_load(&(p -> last_stop_count)) == GC_stop_count) {
            /* The thread has been restarted.   */
            if (!in_resend_restart_signals) {
              /* ... */
            } else {
              continue;
              /* FIXME: Still, an extremely low chance exists that the  */
              /* user signal restarts the thread after the restart      */
              /* signal has been lost (causing sem_timedwait() to fail) */
              /* while retrying, causing finally a mismatch between     */
              /* GC_suspend_ack_sem and n_live_threads.                 */
            }
          }
          n_live_threads++;
          ...

The probability of this is extremely low (probably never observed in practice).

[ivmai]

I think to fix this properly we should replace n_live_threads counter with a flag in GC_thread indicating that sem_[timed]wait() should be called to "compensate" sem_post() call done in the given thread.