fix(vibevoice-asr): add audio resampling and normalization to preprocessing

[bellkjtt]

Summary

• Add audio normalization (-25 dB FS) to match the official VibeVoice PyTorch pipeline
• Add sampling_rate parameter to generate() and stream_transcribe() so that array inputs at non-24 kHz rates are resampled correctly
• Previously, passing a raw waveform array (e.g. loaded via librosa at 16 kHz) resulted in poor transcription because the audio was fed to the model without resampling or normalization

Details

The official VibeVoice ASR pipeline (PyTorch) applies two critical preprocessing steps that were missing in the MLX implementation:

1. Resampling to 24 kHz - The model expects 24 kHz input. When users load audio at a different sample rate and pass it as an array, _preprocess_audio now resamples via resample_audio().
2. Loudness normalization - The official pipeline normalizes audio to -25 dB FS with clipping avoidance. A new _normalize_audio() static method replicates this behavior.

Test plan

• Test with audio loaded at 16 kHz via librosa - should now produce correct transcriptions
• Test with audio file path (str) - should behave identically to before
• Test with audio already at 24 kHz - should apply normalization without resampling
• Compare output quality against the official PyTorch VibeVoice ASR demo

[Blaizzy]

LGTM, thanks!