feat(voice): speak assistant text while it streams

[johnmatthewtennant]

Summary

Implements the conversation side of the duplex stream: streaming LLM response -> TTS.

As ACP text deltas arrive, Berd appends the raw text to the Pocket utterance stream instead of waiting for the assistant message to complete. The backend owns sentence and token-limit segmentation. Tool boundaries flush speakable text without ending the utterance; response completion finishes it.

This is distinct from PR #133, which streams Pocket PCM synthesis chunks to playback and overlaps later synthesis with already-playing audio.

Reviewer-reproducible examples

• Ask for a long multi-paragraph story. The transcript should grow while the model responds, and speech should begin after the first complete sentence, before the full response is visible.
• During that response, watch the text block status. It should show one Speaking state for the utterance, not a separate Speaking block for every synthesized sentence or token chunk.
• Listen across sentence boundaries. Later sentences should already be synthesizing while earlier audio plays, avoiding the long gaps caused by serial synthesize-then-play jobs.
• Interrupt once in the middle of the response. Playback and pending synthesis should stop, with one Interrupted status and one delivery notice for the utterance.