Add macOS Nix and MLX runtime support

[66Ton99]

Summary

• add a Nix-based macOS setup path for Apple Silicon, including MLX runtime dependencies and Web UI build helpers
• add moshi.convert_mlx to create a local MLX bundle, including native model.moshi_mlx.safetensors for moshi_mlx
• route --device mps and auto-selected MPS to the MLX runtime while preserving CUDA/CPU PyTorch paths
• improve local model asset discovery, browser-safe server URLs, voice prompt embedding loading, and Web UI playback buffering for MLX bursty output

Verification

• nix-shell --run './scripts/bootstrap-nix-macos.sh'
• nix-shell --run 'python -m py_compile moshi/moshi/server.py moshi/moshi/offline.py moshi/moshi/convert_mlx.py moshi/moshi/models/lm.py'
• nix-shell --run 'python -m moshi.convert_mlx --self-test'
• ./scripts/build-client.sh
• git diff --check
• MLX smoke test on macOS: python -m moshi.server --ssl "$SSL_DIR" --device auto --port 9016 --static none resolved device = mps, loaded model-mlx/model.moshi_mlx.safetensors, warmed up, listened on https://localhost:9016, and returned WebSocket handshake b'\x00'

It will close #71