docs: MLX backend design (planned)

[wavekat-eason]

Summary

• Adds docs/05-mlx-backend.md — design doc for a future MLX backend that runs the full Qwen3-TTS pipeline natively on Apple Silicon GPU/ANE
• Marked as Planned — not being implemented now; GPU work is moving to CUDA (Google Colab) first
• Documents bridge strategies (mlx-rs, Swift dylib, Swift sidecar), weight format, inference design, and open questions for future reference

Test plan

• Doc renders correctly on GitHub