Refactor the HTTP handler hierarchy to make it simpler.

Replace the ghost logging handler with a simpler one, that fires as soon
as the HTTP request headers are sent (this is a better solution for
long-lived connections).

Add a debug page to track long-lived HTTP request.

Add tests for the HTTP server.

This will limit useless reloads (and related killSources), for instance
when transcoding parameters change. Also, improve the icecast
config generation test.

Adds optional transcoding parameters to a Mount. Each node will then
start a master election for every transcoded stream, and the master will
run liquidsoap to re-encode the data. Transcoding mounts are identical
in other respects to normal mounts, except that liquidsoap acts as a
source for them.

Refactor the load balancing code using a generic traffic control
abstraction that gives the user more control over the final logic (for
the moment this is configured via a command-line parameter).

Adds a new metric to NodeStatus (MaxListeners) so that it is available
to the traffic control logic.

This change allows us to introduce changes to the inner workings of the
cluster that break backwards compatibility, allowing a smooth transition
to the new format.

Increase robustness of the config watcher by reloading the full
configuration on Etcd errors (fixes an issue with the very common error
401).

Make sure that the Icecast config reload is only triggered if the
configuration has actually changed.

Added a way to stop the node properly, which ensures that we are
tracking all the resources and that there are no deadlocks in the
internal communications.

MasterElection now closes the update channel when it's done.