feat(voice): pack natural TTS sentences without fixed silence

[tlongwell-block]

Summary

• preserve the first sentence as the highest-priority Pocket TTS synthesis unit for minimal time to first audio
• replace Desktop's coarse character-based grouping with tokenizer-aware sentence/clause packing
• keep subsequent synthesis pipelined with playback, preserving output order without racing the mutable Pocket engine
• remove the unconditional 100 ms of inter-unit silence so Pocket's generated pauses determine cadence

Why

Buzz currently resets prosody at coarse boundaries, fades each unit, and adds fixed silence even when the model has already generated a natural pause. This makes speech sound more segmented than the Pocket TTS web demo.

This keeps PR #5671's latency strategy: sentence one is isolated and generated first. Once its PCM is queued, the remaining naturally packed units continue synthesizing while playback drains the queue.

This PR does not add parallel ONNX generation. Pocket owns one mutable engine and synthesis remains sequential; true parallel generation requires another resident model/session and should be evaluated separately as an explicit memory/latency tradeoff.

Dependency

Stacked on #5671 (eva/huddle-first-audio-latency) as requested. The merge base is its exact head at the time of implementation: f9451f1f502b12db2c8972edc934d6f2138182b9.

Verification

At 6ba487ecf69e9504505eb0449c3c929e606461d4:

• buzz-voice library tests: 19 passed, 7 model-dependent tests ignored
• full Desktop library tests: 2,412 passed, 15 ignored
• buzz-voice and Desktop clippy clean for all targets with -D warnings
• full pre-push hooks passed using the repo's Hermit toolchain (PATH="$PWD/bin:$PATH")
• regression coverage pins first-sentence playback isolation and verifies the inner model split does not isolate later packed sentences again