Add Apple Silicon MLX inference backend

[ailuntx]

This PR adds an experimental Apple Silicon / MLX inference backend for OmniVoice.

Changes:

• Adds omnivoice.mlx.OmniVoiceMLX with an MLX implementation of the OmniVoice Qwen3-style inference path.
• Adds omnivoice-infer-mlx CLI.
• Adds optional mlx dependency group.
• Adds conversion, quantization, staging, and validation scripts under scripts/.
• Documents local MLX usage in the README.

Validation performed locally on Apple Silicon:

• Official PyTorch/MPS baseline generated short and normal samples from the official checkpoint.
• MLX backend generated short and normal samples from the same official checkpoint.
• Forward-logits alignment vs PyTorch/MPS on a prepared input: shape (1, 8, 17, 1025), identical target-tail argmax tokens, mean abs diff about 0.005.
• Staged MLX variants tested locally: fp32, bf16, 8bit, 4bit.

Related tracking issue: #157

Community runtime/staging mirror: https://github.com/ailuntx/OmniVoice-MLX