feat(cloud): fire estimated word boundaries progressively during streaming

[willwade]

Why

VoiceGarden-SPD (speech-dispatcher module) exposes this: for engines without API timing data (OpenAI, Deepgram, Watson, …), on_boundary estimates fired in one batch after the response completed. Since PR #10 streams audio as it downloads, callers interleaving index marks with playback could only emit them at the end — word highlighting was effectively useless on long utterances for ~15 of the 20 cloud engines.

What

• EstimatePlan — pre-resolves the 150-wpm estimates (char offsets included, SSML stripped first) into firing order
• EstimateFirer — tracks cumulative delivered PCM samples and fires estimate i once ≥ its start-time worth of audio has been emitted. This anchors the text-based estimates onto the real audio clock: a voice slower than the 150-wpm baseline simply holds each mark until its audio arrives; audio shorter than the estimates flushes the remainder at end-of-stream
• StreamEvt — stream_body_to_on_audio now takes a single Audio | Boundary event callback (threading two &mut dyn FnMut callbacks hit trait-object lifetime invariance)
• IncrementalDecoder::sample_rate() — exposes the rate observed in the first decoded packet so MP3 bytes can be converted to seconds
• Raw-PCM bodies (Azure, Cartesia) use their pinned 24 kHz config rate

Real-timing engines (Azure/Edge WS, Google timepoints, ElevenLabs with-timestamps) are unchanged.

Testing

• estimated_boundaries_interleave_with_audio_events: a Boundary event provably fires between Audio events on a dribbled silent-MP3 stream (not at the end flush)
• estimated_boundaries_fire_progressively_during_streaming: every estimate fires; flush covers short audio
• estimate_plan_strips_ssml_before_estimating
• Full cloud suite green (115 tests); clippy -D warnings clean