Thanks to Jeroen van Veen for the debugging.

We used to call it on just the first track.  Also remove obsolete
remains of the onlabel callback.

Avoid formatting JSON multiple times.

Split the id field into id and source, where source indicates the sender
of the message and id the entity being sent.  Remove the label request,
just use the offerer's username.  Maintain the username within the
ServerConnection, this removes a parameter from some methods.

This avoids one HTTP request, and is potentially more flexible.

This could be solved by simply reordering the fields, but it
is more robust to move the atomics into their own structure,
and the extra indirection should not cost much.

We bound the NACK delay to 2 packets, which is what TCP does.
Thus, reducing the delay should be safe.

It is now possible to specify codecs other than VP8 and Opus.  This
turns out not to be very useful, since VP8 is the only codec supported
by all browsers (in violation of the WebRTC spec, which mandates support
for H.264), and there is no good reason to use anything other than Opus
for audio.

This avoids spurious up.complete indications.

This is not an issue in the current code, assuming that clients
negotiate correctly, but will become one once we implement codec
flexibility.

We used to push a connection when we got all tracks.  However, it may
happen that some tracks fail; in that case, the connection would never
get pushed.