It would be better to terminate FIR, note that a keyframe was requested,
and send FIR on our own.

Thanks to Giuseppe Castagna for noticing.

It was racy, and it drops the chat history.  Let them accumulate for now.

This spuriously breaks client-to-server connections over IPv6.
Thanks to Fabien de Montgolfier for his help with debugging this.

It turns out that track ids are not necessarily the same on the local and
remote sides.  Thanks to Ines Klimann for noticing the issue.

The reader and the writer now communicate through a channel and the packet
cache.  If the writer thread drops behind, we drop packets after inserting in
the packet cache, which avoids building a backlog.