Remove a lot of boilerplate by using the errgroup package instead of
our own custom cumbersome solutions. Control server lifetime with an
outer Context.

It is best to just make the DNS server listen on all non-loopback IP
addresses from all interfaces. The loopback exclusion is just to
integrate nicely with an eventual DNS cache running on the same host.

And various other minor fixes.

Handle the case when IcecastMount is nil in the template by using the
protobuf getter.

Fix also some errors in the status template, and update to Bootstrap
4.3.1 and JQuery 3.4.0.

It could be that the "n=0 means EOF" semantics are leaking from the
underlying splice() system call? Anyway this prevents radiod from
getting stuck in a busy loop in certain(?) cases.

It's a float encoded as a string.

Unless an explicit fallback is configured. Untested.

* We can't use a RWMutex due to the predictor write at the end
* Return and log errors when lbv2 returns no results
* Generally increase http debug level

Half-closed connections will get sources stuck.

Drop our own XSLT template (yay), we can get the same information
using the built-in /status-json.xsl page with Icecast >= 2.4.

Use pb.Status as the node information.

At least it doesn't require sudo like using systemctl.

Export the node Status as metrics.

Also add instrumentation for the load balancer module.

We only use the proxy for Icecast streaming so there's no need to be
generic.

Includes a merge of radiod and redirectord, updates to more modern Go
idioms, adoption of the higher level etcd concurrency primitives, and
a lot of other minor implementation changes.

The refactor is still incomplete: the daemons won't compile and the
code to start the RPC servers is missing.

It should work with Debian wheezy (sysvinit) and jessie (systemd)
without having to pass a reload command as a command-line flag.