Audio: add Faster-Whisper ASR adapter

[mohammadaaftabv]

Summary

• add the model-only Faster-Whisper adapter for merged Audio: add Qwen-Omni in-process ASR adapter #1967's generic ASRStage
• use Faster-Whisper's download API for node prefetch, avoiding construction of a full CPU inference model during setup_on_node
• support CPU/GPU model loading, revision pinning, forced-language inference, and Whisper language aliases
• lazily load the optional faster-whisper dependency
• ground inference behavior in nkoluguri/integration-test at 1f1e770b429d47292a4133dd3ab15e46b1cf8ecc

Reused from main

• generic stage resources, worker lifecycle, language allowlisting, notes, skip state, waveform retention/removal, and output assembly
• mono conversion and 16 kHz resampling upstream of the adapter

There is no dedicated Faster-Whisper stage and no adapter-side Torch/torchaudio waveform preparation.

Dependency consistency

faster-whisper>=1.1.0 remains a direct audio_common dependency. It was already resolved transitively through WhisperX; this PR updates root lock metadata so uv lock --check passes independently.

Current diff

• +229 / -0 across 4 files
• lock metadata accounts for +5 lines

Validation

• Ruff passed
• 6 focused tests passed
• uv lock --check passed

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.