An out of file descriptors error was causing the async acceptor loops to
die. The notification process was unaffected so the workers didn't die.
Async workers hitting an error in the acceptor now kill themselves and
rely on the arbiter to restart a new worker in their stead.
Reported by Sergey Shepelev. I changed his patch slightly because of the
man page language. As I read things, the close on exec flag is separate
from the F_SETFL flags.
