Decodes one source to PCM and paces it out as timestamped WireChunks.

ffmpeg is used purely as a decoder: it floods raw s16le PCM into a buffer as fast as it likes. The pacer owns the audio timeline — chunk N is stamped with start + N * chunk_ms (server clock) and emitted once realtime is within bufferMs of that play time. Because the server assigns the timestamps (not arrival order), there is no producer/consumer drift: the only requirement is to send a chunk before its play deadline, and the bufferMs lead absorbs all jitter.

source is a binary path/URL, or a 0-arity function returning one (resolved when the stream starts, e.g. for short-lived signed URLs).