It's worth testing that our transition to the DEAD state works as
expected, by intentionally killing a server.

This test also makes it possible to test what happens when pending
commands are stranded, so that an upcoming patch to add notifiers can
show the difference it makes.

The test was surprisingly hard to write. For starters, sending SIGINT
to nbdkit was sometimes enough to kill the process (if it hadn't yet
read the NBD_CMD_READ, and therefore did not try to wait for any
outstanding requests before quitting), but often it did not (because
nbdkit was stuck waiting for pthread_join()). Then there's the race
that nbd_poll() can sometimes get lucky enough to catch a POLLHUP in
REPLY.START where recv() returning 0 transitions things to CLOSED, but
more often catches a POLLERR and transitions to DEAD.  Then there was
the hour I spent scratching my head why kill didn't seem to get rid of
the child process even though the poll() was definitely seeing the fd
closing, until I remembered that kill(pid,0) to zombie processes
succeeds until you wait() or alter SIGCHLD to specifically prevent
zombies.  And debugging the now-fixed uninitialized variable in
nbd_unlocked_poll added to the mix.

I'm not 100% sure the test is portable to non-Linux - I guess we'll
eventually find out when someone worries about a BSD port of libnbd.