audio: add local Whisper backend via candle (feature-gated)

[bug-ops]

Parent: #520
Depends on: #522

Context

For offline/air-gapped environments and zero API cost, provide local speech-to-text using Whisper via candle. Feature-gated behind `candle` (or a new `whisper-local` flag).

Design

Implementation

Leverage candle-transformers Whisper example as reference:

• Download model weights from HuggingFace (whisper-tiny, whisper-base, whisper-small)
• Audio decoding: use `symphonia` crate for format handling (ogg, mp3, wav, flac)
• Mel spectrogram computation + whisper inference via candle

Model options

Model	Size	VRAM	Quality
whisper-tiny	39 MB	~1 GB	Good for short commands
whisper-base	74 MB	~1 GB	Better accuracy
whisper-small	244 MB	~2 GB	Best quality, still fast

Config

[audio]
backend = "whisper-local"

[audio.whisper_local]
model = "whisper-base"  # tiny, base, small
device = "auto"         # cpu, metal, cuda

Feature gate

[features]
whisper-local = ["dep:symphonia", "dep:candle-core", "dep:candle-nn", "dep:candle-transformers", "dep:hf-hub"]

Acceptance criteria

• Implements `SpeechToText` trait
• Auto-downloads model on first use via hf-hub
• Supports Metal (macOS) and CUDA (Linux) acceleration
• CPU fallback works
• Feature-gated, does not affect default build
• Transcription of 10s audio completes in <2s on Metal