Rewrite Local Wisper primarily in BAML

[none23]

Summary

Replace the Python application with a compiled BAML application for local Parakeet speech-to-text. The existing lw command and Sway keybinding workflow remain compatible. The installed app runs without Python or the BAML CLI and automatically chooses CUDA or CPU inference.

Changes

• Run nvidia/parakeet-tdt-0.6b-v3 through native ONNX inference instead of Python, NeMo, or Faster Whisper.
• Prefer FP16 CUDA inference and fall back to INT8 CPU inference when CUDA is unavailable. Keep one warm model process and one verified model cache per user to avoid duplicate RAM or VRAM use.
• Preserve the lw, preload, sway-start, sway-stop, and sway-cancel commands, the existing Sway wrapper contract, and clipboard or focused-window transcript delivery.
• Preserve deterministic glossary and spoken-number cleanup, with optional OpenAI cleanup for longer transcripts.
• Treat speech-to-text as a system-level feature rather than an editor feature, and remove the Neovim plugin.
• Remove user-facing backend, model, device, and quantization choices in favor of one automatic configuration.
• Add automated BAML, Rust, and shell quality checks for contributors and pull requests.