feat: stream synthesis sentence-by-sentence for low-latency playback

[StuBehan]

Summary

Makes speak() stream synthesis sentence-by-sentence so audio starts almost immediately instead of after the whole text is synthesized. Kokoro batches by phoneme count (~510), so a full paragraph synthesizes as one blob before any sound plays — that's the lead-in lag. Splitting into sentences ourselves means the first sentence (~0.2s to synthesize) plays right away while the rest synthesize in the background.

Changes

• _split_sentences() — breaks text into sentence-ish chunks on ./!/? followed by whitespace, and on newlines. The whitespace requirement leaves decimals (0.95) and mid-token dots (file.ts) intact.
• _stream_play() — a producer thread synthesizes ahead into a bounded queue (maxsize=8) while the main thread plays each chunk as it's ready, so later sentences are usually synthesized by the time the previous one finishes.
• speak(blocking=True) streams inline; blocking=False runs _stream_play on a background daemon thread (_play_thread).
• stop() now sets a _stop_event so an in-progress stream halts between sentences (in addition to sd.stop()).
• Per-call voice/speed/lang overrides are threaded through to each sentence's synthesis.

Testing

Ran the full CI-equivalent suite locally, all green:

ruff check .            # All checks passed!
ruff format --check .   # 18 files already formatted
mypy                    # Success: no issues found in 16 source files
pytest -q               # 126 passed

New tests cover: one synth+play cycle per sentence, single-sentence playback, background-thread playback for blocking=False, engine defaults vs per-call overrides, stop() setting the event, and _split_sentences edge cases (decimals/mid-token dots preserved, newline splitting, blank input).

Related issues