Use notification pipe hangup as shutdown event notification response.

This now handles POLLHUP and POLLNVAL on the shutdown command response
pipe as shutdown notifications, closing a race where the client thread
called poll(2) just after the sender thread closed the notification
pipe, leading to an unnoticed POLLNVAL & retry cycle, blocking forever.
(This race condition was easiest to reproduce by setting up a
KineticClient, then shutting it down without doing anything, in a loop.)

Also, add some shutdown logging statements (with high log levels) and
make sender_register_socket and sender_remove_socket calls ignored
after sender_shutdown has been called.