Some codecs (VP9) can both buffer _many_ frames, and take a long time
encoding each frame. Accumulated, the last `encode(None)`-flush can end
taking a long time >30s, without detectable progress.

FFmpeg and many encoders themselves output one frame at a time, but PyAV
currently buffer them all up into lists returned.

This change adds a `encode_lazy` yielding frames as they are made ready.

The change was benchmarked to also yield a net performance improvement. For
both `encode()` and `encode_lazy` encoding really small (24x18) frames
using the `mpeg4` encoder seem to take ~11% less time.