libflux/ev_flux: handle EV_ERROR on io watcher

Problem: ev_flux ignores events on its internal io watcher,
so when libev raises EV_ERROR on it because something went wrong
internally (like flux-framework#2554), nothing happens.

The internal io watcher is only used for its side effect of
unblocking the reactor when pollevents edge triggers from
"no events" to "some events".  The prep/check/idle watchers
do the heavy lifting.

Add a check for pending EV_ERROR events on the io watcher in
the check callback.  If found, notify the user via the ev_flux
watcher callback.

Note: in libev 4.25, a failure of epoll_ctl() would internally
call fd_kill(), which would call ev_feed_event() to raise
EV_ERROR on the watcher.  With this patch, EV_ERROR is caught;
however, only after a one minute delay, which has not been
explained.  libev 4.27 turns that error into an assertion.

Since other error paths in libev call fd_kill(), this change may
be useful for other as yet unseen problems.

src/common/libflux/ev_flux.c

@@ -49,6 +49,12 @@ static void check_cb (struct ev_loop *loop, ev_check *w, int revents)
 {
     struct ev_flux *fw = (struct ev_flux *)((char *)w
                             - offsetof (struct ev_flux, check_w));
+
+    if (ev_is_pending (&fw->io_w)
+            && ev_clear_pending (loop, &fw->io_w) & EV_ERROR) {
+        fw->cb (loop, fw, EV_ERROR);
+        return;
+    }
     int events = get_pollevents (fw->h);
 
     ev_io_stop (loop, &fw->io_w);