Do you pipe frames directly through a bounded queue to keep ffmpeg decoding from outpacing the embedding model?