This avoids allocating/freeing an AVFrame for each packet decode. It also
simplifies error handling in transcode_audio() and transcode_video().