We need support three distinct paths for destruction of
ncsubprocs: (1) external call to ncsubproc_destroy() while
the subprocess is running, (2) external call to
ncsubproc_destroy() after the subprocess has terminated, and
(3) internal call to ncsubproc_destroy() without any external
call. To do this properly, we always waitid() on the subprocess
in our ncsubproc thread, and do not cancel said thread. This
guarantees that the subprocess has been reaped if the thread
has exited. We throw a pidfd_send_signal() into the thread
prior to the waitid(), because this is safe with pidfds. The
thread reclaims no resources otherwise. ncsubproc_destroy(),
instead, reclaims them, after joining the ncsubproc thread.
It sends SIGKILL before the join, which is once again safe
thanks to pidfds. Resolves#552.
