Run Qwen3-ASR natively in Swift

[IchenDEV]

Summary

• Replace the Qwen3-ASR Python virtual environment and subprocess runner with in-process Swift/MLX inference through mlx-audio-swift 0.1.3.
• Reuse the existing mlx-community/Qwen3-ASR-1.7B-bf16 model location and keep model downloads under the existing model catalog.
• Remove the bundled Python runner and app-managed Python runtime code.
• Keep MiMo unavailable in the released UI and migrate legacy MiMo selections to Apple Speech.
• Leave the existing GitHub Release, self-signing, DMG, Info.plist, and entitlements flows unchanged.

Verification

• bash scripts/ci-basic-checks.sh
• DEVELOPER_DIR=/Applications/Xcode-beta.app/Contents/Developer swift test
  • 557 XCTest cases passed, 8 environment-gated tests skipped
  • 1 Swift Testing case passed
• Offline existing-model integration with HF_HUB_OFFLINE=1
  • English: 2.034 s, expected design-doc transcript
  • Chinese: 0.439 s, expected 周五 transcript
  • model weight size, modification time, and SafeTensors set unchanged
• scripts/build-app.sh --app-only --sign=-
  • generated an arm64 ad-hoc signed app
  • codesign --verify --deep --strict passed
  • bundled default.metallib
  • no Python files, Python runner, pip package marker, or runtime-loaded .so

Boundaries

• The native ASR validation currently covers short English and Chinese repository samples; long audio, repeated stress runs, and mid-generation cancellation still need broader product validation.
• MiMo remains intentionally disabled.
• This change does not modify or trigger the tag-based GitHub Release workflow.