Instead of buffering the last keyframe, we merely keep track of its
seqno, and use the main cache for recovering.  We also send the whole
sequence of packets rather than just the keyframe itself.

Rewriting is not useful for VP9, and even harmful, as it breaks the
offsets in the scalability structure.

If the client requested a low-resolution stream and there is none,
switch to SID=0.

The sync bits were incorrect.  In addition, we need to set the
marker at the new end of the frame when doing spatial decimation.

Otherwise we might act on outdated info, which could cause a spurious
drop.

Left over from the old data structure.

Now that we constrain tracks in offers properly, there is no need
to manually check that the codecs have been matched correctly.

We now restrict the allowable codecs in the downstream direction,
which leads to a clean failure instead of a silent track.  We also
allow multiple profiles for a single codec.

The lower interval made us too sensible too scheduling.

Only used as a last-resort fallback.

Pion is making incompatible changes, so stick with webrtc 3.0 and
rtp 1.6 for now.

We used to check the nick, which is not necessarily unique.

This replaces the OverridePermissions method, and is communicated
to clients that can reliably and safely identify a system user.

The keyframe detection has been moved into codecs.

We store h.264 in Matroska, since we already have the library.
We currently store the video width and height as 0.

Instead of testing whether a sample is a keyframe, we test at the
packet level, then compare timestamps to identify the keyframe.

The trick of calling the onclose callback wasn't correct.

This fixes issues introduced in commit 6ae79f.

We now inform clients of the status of a group (locked, etc.).  Also
cleans up the handling of administrative messages, which solves the
issue of receiving "user" before "joined".