feat(tts): add MiniMax Music 3 inference

**Because**

- mlx-audio could not convert or run the official MiniMax Music 3 modular checkpoint, leaving Apple Silicon users without native lyric-conditioned song generation.
- Structural loading alone was insufficient evidence for a large multi-stage port, and the lower-memory MXFP4 path needed honest quality guidance.

**This commit**

- Adds the native Qwen3/RVQ autoregressive stage, condition encoder, flow-matching DiT/Euler pipeline, and stereo 44.1 kHz vocoder.
- Extends conversion for modular checkpoints and produces strict-loadable dense, MXFP4, and MXFP8 layouts while keeping sensitive heads, convolutions, and the vocoder dense.
- Registers lyrics, duration, steps, and seed controls; documents MXFP8 as the recommended quantized format and MXFP4 as an experimental memory tradeoff.
- Preserves Apache-2.0 attribution and packages LICENSE/NOTICE files with the wheel and sdist.
- Verifies 655 TTS/registry tests, three official PyTorch differential parity tests, strict docs, package contents, and dense/quantized generation auditions.

Co-Authored-By: Codex <noreply@openai.com>