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.