Discard frames where the size does not match the AVCodecContext width/height. This avoids possible crashes on video size changes. The problem is that we reinitialize the vo on get_buffer but due to codec delay libavcodec might still return frames with the old size afterwards, which the vo might no longer be able to handle. Ideally libavcodec should not show this behaviour, since it requires that any application using DR1 can handle frames of different sizes simultaneously - which seems a bit extreme.