Add downloadable local speech models and model comparisons

[note89]

Summary

• Add in-app download, cancellation, removal and selection for Cohere Transcribe, Qwen3-ASR, Parakeet Redux and Whisper large-v3-turbo. Verify pinned packages before installing and run native MLX/Core ML inference entirely locally.
• Compare download sizes, measured loading plus first transcription, repeat-dictation measurement status and every supported language with flags and names. Keep the selected model loaded across dictations.
• Share the transcription and cleanup pipeline with a repeatable recorded benchmark that reports startup, latency, accuracy and cleanup fallback, with private local corpora and baseline comparisons.

Test plan

• Swift test suite: 60 tests in 9 suites, including package verification/cancellation, audio conversion, chunk coverage, supported languages and benchmark scheduling.
• Independent implementation reviews; strict Swift formatting/lint, shell syntax and property-list validation.
• Download and transcribe with all four models while networking is denied; exercise Swedish/Portuguese where supported and a 462.89-second Cohere recording.
• Bake-off scoring checks and signed hardened-runtime app bundle verification.
• macOS PR CI: Swift tests, Bake-off scoring, release build and hardened-runtime bundle verification.

Performance values are initial cold synthetic-speech measurements on an M3 Pro with 18 GB memory; warm dictation latency and representative accuracy remain unmeasured. Interactive Download-button clicking has not been verified.