Model request + implementation plan: dots.tts (rednote-hilab) zero-shot voice cloning

I'd like to add dots.tts zero-shot voice-cloning support using the Apache-2.0 MLX runtime from sb1992/dots-tts-mlx. The implementation would vendor the runtime under `mlx_audio/tts/models/dots_tts`, replace its mlx-lm imports with `mlx_audio.lm`, and expose mlx-audio's standard `Model`/`ModelConfig` and `GenerationResult` interface. Because the checkpoint contains three independently loaded safetensors, I propose adding a small generic loader hook letting self-loading models skip the normal merged-weight pass instead of relying only on `sanitize() -> {}`. I have a working branch (black/isort clean, 673 tests passing incl. 22 new, full Apache-2.0 notice packaged for the vendored work, validated end-to-end tonight against the published community int4 weights) and before opening the PR I'll also provide a root-loadable model repo path or subfolder mechanism (the published weights live in variant subdirectories), a real generation smoke test, and cold-load memory measurements. Would you prefer this vendored model package per house convention, or an optional adapter around the external dots-tts-mlx package?