feat: add CosyVoice3 MLX implementation

[walkoncross]

Summary

Adds a full MLX port of CosyVoice3 (mlx_audio/tts/models/cosyvoice3/), following the existing model contract used by other TTS models in this repo (e.g. chatterbox, spark).

• Pipeline: text → LLM (Qwen2 speech-token backbone) → speech tokens → flow (DiT-based CausalMaskedDiffWithDiT / CFM) → mel → HiFT vocoder (causal) → waveform.
• Generation modes: zero-shot voice cloning, cross-lingual synthesis, and instruction-based (instruct2) generation, matching the reference CosyVoice3 Python implementation's behavior.
• Speaker presets: spk2info system for caching/reusing zero-shot speaker prompts (--add_spk, --save_spkinfo, --load_spkinfo).
• Text frontend: ported text normalization pipeline (Chinese blank/bracket handling, English number spelling) and full CosyVoice3Tokenizer special-token registration for fine-grained control tokens.
• Weight conversion: sanitize/routing logic wired into convert.py so official CosyVoice3 checkpoints can be converted to MLX weights.
• Correctness fixes along the way (each validated against the PyTorch reference under /CosyVoice/cosyvoice): causal HiFT rewrite, DiT RoPE pairing/application order, CFM noise seeding to remove pitch bias, sample-rate handling for timbre cloning, prompt-feature/silence-token alignment with the reference, and a defensive post-hoc fix for a rare stochastic LLM-sampling tail-hiss artifact (acoustic-signature detection + fade, since the exact triggering token sequence could not be reliably reproduced after 100+ attempts — see README.md under this model's ## Status section for the full investigation).
• Docs: CLI usage examples (zero-shot, cross-lingual, instruct2, speaker presets) with bundled reference audio assets, and a ## Status section tracking known issues/investigations.

45 commits, squash-mergeable if preferred. See individual commit messages for a detailed history of the implementation and each fix.

Test plan

• pytest mlx_audio/tts/tests/test_cosyvoice3.py — 31 passed (2 pre-existing failures unrelated to this branch, confirmed via git stash against main: TestLLM::test_autoregressive_decode, TestModelIntegration::test_generate_end_to_end_with_mock_frontend, both AttributeError on an incomplete MockFrontEnd test double)
• End-to-end generate() validated against real CosyVoice3-0.5B checkpoint weights for zero-shot, cross-lingual, and instruct2 modes
• HiFT/flow determinism verified (fixed mel → bit-identical waveform across repeated calls)
• Tail-hiss fix verified: real reported bug file -28.1dB → -120dB after fade; synthetic clean-tail audio left bit-identical (no false positives)

[lucasnewman]

Can you remove the implementation-specific comments here and elsewhere? We only want comments that improve the understanding of the code, not porting notes.

[lucasnewman]

Can we just use make_sampler here or is there something specific needed for these?

[lucasnewman]

These should be passed through to the model as extra kwargs, we don't want these in the standard generate interface.

[lucasnewman]

Please revert these.

[lucasnewman]

@walkoncross Thanks for the submission! This needs some cleanup, I left a few comments inline. Also, the examples should just live in the README.md, we don't want model-specific artifacts in there. If you can make those changes we can consider this.

[lucasnewman]

See comments above.