feat(voice): stream Pocket audio during synthesis

[johnmatthewtennant]

Summary

Pocket TTS streams synthesized audio chunks into a single player as decoder output becomes available. Synthesis and playback run concurrently within one utterance.

The backend accepts incremental text, releases the first complete sentence immediately, and forms later model-safe chunks within Pocket's 50-token input limit. Cancellation and delivery status remain scoped to the whole utterance rather than its internal text or audio chunks.

This is audio-output streaming from Pocket into playback. Streaming an LLM response into the TTS input is a separate path.

Reviewer-reproducible examples

• Play a long Pocket-voiced response. Confirm playback begins after the first text chunk is available while later chunks are synthesized concurrently.
• Interrupt a long response shortly after playback starts. Confirm current audio and queued synthesis stop together and no later audio resumes.
• Observe the response status during playback. Confirm the text block receives one speaking, spoken, or interrupted lifecycle rather than one status per internal synthesis chunk.

The merge-base changed after approval.