Patch release for the transcription model picker and downloader.
- Refresh downloaded model state immediately while deleting local model files.
- Start model downloads directly from the transcription settings picker.
- Clean up model display names, NVIDIA row spacing, and picker action pill sizing.
- Show selected and warming model state directly on the relevant picker row instead of a separate warming row.
- Treat leftover aria2 sidecar files as incomplete Voxtral downloads so partial safetensors caches are not marked usable.
- Use the current macOS runner Xcode toolchain and skip package plugin validation for release builds.
