Kill misbehaving subscribed clients instead of hanging

This change only affects clients that are subscribed to events, which
should be the main cause of our problems.

In the common case (no buffered data) the behaviour doesn't change at
all: the message is sent directly, no ev_io / ev_timeout callback is
enabled. Once a write to a client's socket is not completed fully
(returns with EAGAIN error), we put the message in the tail of a queue
and init an ev_io callback and a corresponding timer. If the timer is
triggered first, the socket is closed and the client connection is
removed. If the socket becomes writeable before the timeout we either
reset the timer if we couldn't push all the buffered data or completely
remove it if everything was pushed.

We could also replace ipc_send_message() for all client connections in
i3, not just those subscribed to events.

Furthermore, we could limit the amount of messages stored and increase
the timeout (or use multiple timeouts): eg it's ok if a client is not
reading for 10 seconds and we are only holding 5KB of messages for them
but it is not ok if they are inactive for 5 seconds and we have 30MB of
messages held.

Closes #2999
Closes #2539

include/libi3.h

@@ -166,6 +166,14 @@ int sasprintf(char **strp, const char *fmt, ...);
  */
 ssize_t writeall(int fd, const void *buf, size_t count);
 
+/**
+ * Like writeall, but instead of retrying upon EAGAIN (returned when a write
+ * would block), the function stops and returns the total number of bytes
+ * written so far.
+ *
+ */
+ssize_t writeall_nonblock(int fd, const void *buf, size_t count);
+
 /**
  * Safe-wrapper around writeall which exits if it returns -1 (meaning that
  * write failed)