Ask FFmpeg to seek to the next packet boundary *before* the seek
position, so we don't miss audio data.  Now we get too much, but we'll
solve that in the next commit.