The test was buggy.

This reverts commit 2eca0e44.

We used to simply synchronise on the first keyframe.  This
could lead to important time offets, since every audio frame
is a keyframe, while we only request video keyframes every
four seconds.

Fix that by parsing the RTCP synchronisation data and adjusting
timestamp offsets accordingly.  If we have no RTCP sync, we simply
assume that the first frames received are synchronised.

Only call close if there's already a writer, which avoids flushing
packets before the first call.

We used to set high-quality audio by default for screenshare,
but that makes the video blurry under bad network conditions.
Obey the hqaudio setting for all streams.

We used to maintain stream information for all users except oneself.

This makes commands such as /set work before you connect.

This is no longer necessary since we use "pre-wrap" in the CSS.

The distinction is not likely to be useful, and it doesn't make
sense in all cases.

This may be useful to get around network blocks in some cases.

Some devices only support the Constrained Baseline profile.

This must be larger than the samplebuilder's MaxLate.

It turns out that the bandwidth limitation is not effective
in Firefox when the rid is set unless we're simulcasting.

Firefox doesn't include the rid in the statistics result.
Use the SSID instead.

Firefox uses a non-standard protocol for simulcast, which has been
supported by Pion for a while now.  Enable simulcast in Firefox.

Thanks to Miroslav36.

Our tokens are not revocable, they should not be generated with
a large lifetime.  Don't encourage users to do that.