Probably not necessary, but I'm supersticious.

We already send NACKs when a packet is missing.  Under high packet loss,
however, the recovery packet might get lost two.  Forward receiver NACKs
to the sender, but only after a delay and after checking that the packet
has not arrived in the meantime.

Avoids seqno wraparound.  Unlikely, but I'm supersticious.

This avoids losing packets at the beginning of a connection.

This fixes corruption of the first keyframe of every stream.

This avoids sending NACKs for packets in the future.

Now that we cache keyframes, it is worth keeping some history
even when latency is low.

Keeps the race detector from complaining.

This should fix the deadlock on pc.Close that we have been seeing
(and that was worked around in a previous commit).

Apparently Close can take unbounded amounts of time.

This may happen if we receive an answer for a connection that has been
closed in the meantime.