feat(voice): add voice input with local whisper transcription and claude cleanup

[tyulyukov]

Summary

• Server-side voice transcription: Added Effect-based TranscriptionService with local Whisper model support via HuggingFace transformers
• Child process architecture: Whisper inference runs in a forked child process to avoid blocking the main thread during model loading/inference
• Model management: Automatic installation of Whisper models (tiny/base/small/medium) on first use, cached in ~/.marcode/whisper-models/
• LLM cleanup: Optional Claude-powered cleanup of raw transcription (removes filler words, fixes grammar while preserving technical terms)
• Voice settings: New VoiceSettingsSection in settings for model selection, language choice, and cleanup toggle
• Web integration: Voice recording hook (useVoiceRecording), mic button in composer, keybinding mod+shift+v to toggle recording
• WebSocket API: New methods for transcription, cleanup, model install/delete; push channel for download progress
• Persistent context: Additional directories now removed from attachments popover (refactored to DirectoryPickerPopover in composer footer)

Testing

• Run bun fmt, bun lint, bun typecheck to verify code quality
• Voice recording toggle via mod+shift+v keybinding
• Test transcription with audio input in different languages (if enabled)
• Verify model installation flow and progress updates
• Test cleanup toggle with and without Claude cleanup enabled
• Verify additional directories picker still works in composer footer
• Check WebSocket test expectations updated for new whisper field in server config