This is an attempt to fix a problem where sources would become stuck
on master -> slave state transitions.

This allows running the service on a single port: with the
--enable-icecast-proxy option, all requests can go through
redirectord.

The etcd server now returns the current index on a failed Create()
call, so we can use that to start the slave Watch.

Fixes a bug where master changes that did not involve the
local node were not triggering updates.

Make progress anyway when the history is gone.

go-etcd now closes the channel passed to Watch() on error, so we need
to restart the syncer every time and use a temp channel.