feat(voice): stream Pocket audio during synthesis

[johnmatthewtennant]

Summary

Implements the Pocket side of the duplex stream: streaming TTS synthesis chunks -> playback.

Berd now owns one Pocket engine and one audio player for an utterance. The backend accepts incremental text, releases the first complete sentence immediately, then forms model-safe chunks under the 50-token Pocket limit. Each chunk is synthesized while already-produced PCM continues playing, and decoder PCM blocks are appended to the same player as soon as they are available. Synthesis no longer waits for prior playback to finish.

This PR provides the backend stream. PR #135 connects streaming LLM text to it.

Reviewer-reproducible examples

• With a long Pocket-voiced response, listen at the boundary between the first and second sentences. Speech should continue without the old synthesis-sized silence between them.
• Interrupt a long response shortly after playback starts. Audio and queued synthesis should stop as one utterance, rather than advancing through separately tracked sentence jobs.
• Observe the response status during playback. The backend emits one started/completed or started/interrupted lifecycle for the whole utterance, not one lifecycle per synthesis chunk.