* This option defaults to true:  if there are no default servers and a watcher
  reports no backends, then use the previous backends that we already know
  about.
* Factor out code that updates '@backends' into the watcher base class.

If the configuration specifies a file_output key, synapse will manifest
and update registrations on the filesystem in an atomic way.
This is useful for applications that do not wish to communicate with
service backends via haproxy

Note that at this time only the Yelp fork of haproxy has support for
redispatching on every retry.

This reverts commit 136d1fe.

We can allow socket updates to go through any time that the config was
updated, but restarts must be rate limited for stability

We are experiencing some very slow updates in production and I think it
may be due to the connection pooling, try :per_callback threading to see
if that helps.

Also fixes default values for do_writes, do_reloads, do_socket

This fixes a bug where we could have a session expiry but not fail pings
because the pooling code would not actually tear down the connection

Adds a comment about when we write out the file output and restored
random backend ordering.

Added documentation of the state_file cache options as well as a few
comments. Also fixed some error messages to be more clear.