…s re-export)

…arams into model config

Frame geometry (sample rate, window/hop, FFT size, pre-emphasis, min frames)
is FSMN-specific and now lives on VADModel.featureConfig (supplied by the models
registry) instead of hardcoded constants in the task. The pipeline and streamer
are parameterized by it; detection thresholds remain generic VADOptions.

Framing (windowing, mean-removal, pre-emphasis, Hann) was ~86% of a detect()
call on device — ~40ms of Hermes vs ~6ms for the model forward pass. Port it to
a native `speech.frameWaveform` op that writes directly into the pre-allocated
model-input tensor, fusing mean-removal + pre-emphasis + Hann into a single
dependency-free (vectorizable) pass. Framing drops to ~3.4ms on device (~12x),
leaving it below the model's own inference cost.

The streamer re-ran the model over the whole growing buffer every tick
(up to ~8ms on a 10s buffer on device), even though only ~100ms is new
and the streaming decision only needs recent context. Cap the sliding
window to 2.5s so per-tick detect stays flat and low (~3x cheaper, ~2.7ms)
— still well above FSMN's receptive field and the 250ms min-speech-duration,
so detection is unaffected. On-device benchmark showed the alternative
'O(1) insert' buffer only saved ~6us/tick (0.2%), so it was dropped.

…ng, op API

- Replace the streaming generator with push()/resetStream(): a rolling window
  fed straight from a recorder callback, returning speechStart/speechEnd. The
  window is required — one 100ms chunk yields ~7 frames, far below the 25 hops
  minSpeechDurationMs needs, so detect() on a bare chunk never fires.
- Rename to createFsmnVoiceActivityDetector / useVoiceActivityDetection, mirroring
  createWhisperSpeechToText / useSpeechToText; frameWaveform -> extractFrames.
- Case acronyms per repo convention (cf. NmsOptions): VADOptions -> VadOptions etc.
- Drop VadFeatureConfig: model geometry is now file consts (fftLength read from
  model metadata) and the sample rate is exported as FSMN_VAD_SAMPLE_RATE_HZ,
  mirroring WHISPER_SAMPLE_RATE_HZ.
- extractFrames takes an options object and no startSample; the caller passes the
  per-chunk waveform slice, which removes the nested try.
- Fold scoresToSegments/mergeSegments into a single postprocess().

The FSMN model's output is dynamic, not just its input: N input frames
produce [1, N, classes], so execute() validates outputTensors against the
shape produced for the current chunk rather than the model-declared maximum.
The output tensor was pre-allocated once at that maximum, so every execute
threw 'outputTensors[0] must have shape [1, 247, 248] ... got 1000'. Inside
the recorder callback nothing surfaced the throw, so live detection silently
never fired.

Allocate the output per chunk with its frame dimension matched to the chunk
and dispose it alongside the other per-chunk tensors. Verified on a physical
Android device: live mic VAD toggles SPEAKING/SILENT again.

- Inference method detect/detectWorklet -> detectVoice/detectVoiceWorklet
  (CV uses task-specific names: classify, detectObjects, transferStyle...).
- Hook useVoiceActivityDetection -> useVoiceActivityDetector (agent-noun,
  matching createFsmnVoiceActivityDetector and the other use* hooks).
- Registry key models.vad -> models.voiceActivityDetection (all other keys
  use full names; the 'vad' on-demand feature flag is unchanged).

…models registry