Feat: Support for Qwen3-ASR 0.6B

**Is your feature request related to a problem? Please describe.**
When dictating technical jargon (e.g., saying `mod tests` or syntax-specific terms), purely phonetic models (like Parakeet) output conversational approximations rather than correct programming terms.

**Describe the solution you'd like**
Add local support for **Qwen3-ASR 0.6B** via MLX. Because it relies on a LALM architecture with an integrated LLM decoder, it can use surrounding semantic context to resolve acoustic ambiguity and handle codespeak significantly more accurately.

**Describe alternatives you've considered**

* **Parakeet TDT v3:** Extremely fast, but lacks the contextual language intuition needed to format non-generic terms properly.
* **Whisper Large v3 Turbo:** Highly accurate vocabulary depth, but has a higher memory footprint and slower execution speed. Not ideal for local STT imo.

**Additional context**
When quantized to 4-bit via MLX, the 0.6B model requires only ~1.2 GB of unified memory and achieves RTF of ~0.06. This allows it to run properly on base Apple Silicon configurations (like the fanless MacBook Air) without thermal throttling.

No files, tests, or entry points are identified. Start by locating the existing local transcription integration and determining where an MLX-backed model would fit; done means Qwen3-ASR 0.6B runs locally on supported Apple Silicon and improves technical-term transcription as described.

Written by the indexing model from the issue text.