feat(sherpaonnx): stream audio per sentence batch via the generate callback

[willwade]

Why

The README claimed sherpa-onnx "synthesises whole clips by design" — that was wrong. Verified in sherpa-onnx's C++ sources (offline-tts-vits-impl.h / kokoro / matcha / supertonic): the generate progress callback receives each sentence batch's newly generated samples as synthesis progresses (the Rust crate's "samples generated so far" doc comment was the error). We already set max_num_sentences=1, so batches are sentence-sized. js-tts-wrapper's synthToBytestream looked like streaming but enqueues one whole-clip buffer — this does the real thing.

What

• Sentence-level streaming: speak() emits each batch through on_audio immediately; volume/pitch applied per batch (batches are sentence-aligned, so per-batch pitch resampling never seams mid-speech). Single-sentence utterances still complete before delivery (inherent).
• Progressive estimated boundaries: shares the EstimateFirer machinery introduced in feat(cloud): fire estimated word boundaries progressively during streaming #14, with a 1/speed time-scale so estimates anchor to delivered audio. Reported times stay on the rate-1.0 baseline — existing rate-compensating callers (VoiceGarden-SPD, the SAPI adapter) are unaffected.
• New boundaries module: EstimatePlan/EstimateFirer moved out of cloud_engine into an always-compiled module; the firer now owns its plan so it can move into 'static callbacks.
• 'static callback plumbing: the generate callback can't capture method-lifetime borrows, so on_audio/on_boundary are stashed as lifetime-erased pointers in thread-locals for the duration of the synchronous call (same-thread invocation per the C++ docs; serialised by the tts_instance mutex — same pattern as VISEME_CB), and the firer is shared via Arc<Mutex<..>> so the outer scope flushes remaining estimates after generation.

Testing

• New live test sherpa_streams_audio_per_sentence_batch (piper-nl-rdh-low): asserts a Boundary event fires before the final audio chunk and first-audio < last-audio — both hold with the real model
• Full local suite green (294 tests incl. moved boundary tests); clippy -D warnings clean; live vits/kokoro families pass (matcha tests skipped locally — model not downloaded, CI covers)
• README streaming notes corrected