convert Mosh from using main loops to event-driven code

Currently mosh-client, mosh-server and the test programs use main loops with a select() at the top, and selective invocation of various functions depending on timeouts or ready file descriptors.

There's a few problems with this:

• The handling of timeouts for the network code is a bit tangled because of the lack of explicit timeout events
• mosh-client does blocking writes to the user's terminal, which can distort time-dependent calculations if the terminal blocks. The is probably also true for mosh-server and writing input to the pty to the user's session. There's other parts of Mosh where timers and the current timestamp can get disconnected from reality, too.
• Adding new files and/or events (like ssh authentication) to Mosh results in more complexity

Converting Mosh to use events, probably using something like libevent, would probably clarify Mosh's internal structure significantly. I don't think it would be too difficult, but it's not trivial either. Things that would need to be done:

• In mosh-client, the terminal display would need to run asynchronously from the network code, writing to the terminal in possibly multiple non-blocking writes. I think this means that the terminal display would have to remember (and retain in saved_states) the last state that was written to the terminal, and write the diff between that state and the most-current state on the next go around.
• Network timers would need to be significantly reworked to use timers and events independent of code execution paths

Start by mapping the select()-based main loops in mosh-client, mosh-server, and the test programs, then evaluate whether libevent fits the proposed event model. Focus on asynchronous terminal and pty writes, retained saved_states, and independent network timers. Done means the main loops use explicit events without the described blocking and timeout problems.