optimize ace_step: completely remove torch runtime dependency + unified conversion pipeline

[mm65x]

Description

this PR builds directly on the massive native MLX implementation from #499, optimizing the ace_step implementation by completely removing the need for PyTorch at runtime, and adding a robust unified conversion script.

i noticed two small things missing from the branch that this PR addresses:

1. the ace_step generation module still fell back to import torch just to load the silence_latent.pt tensor, which technically means users still need a massive PyTorch installation just to run inference.
2. there wasn't a standard convert.py script for users to pull the raw HuggingFace weights and quantize/package them into a neat safetensors directory format locally.

Changes in the codebase

• created a robust convert.py that hooks directly into the upstream HuggingFace repository to extract the DiT, the VAE, and the Text Encoder.
• the conversion script intercepts the PyTorch meta tensors and automatically converts the silence_latent.pt into a pure silence_latent.npy numpy array.
• removed the import torch runtime dependency in ace_step.py completely, hooking it up to read the npy file via numpy instead!

this makes the generation pipeline 100% native PyTorch-free at runtime!

[lucasnewman]

@mm65x Any change you can put the converted model in mlx-community on the HF hub? It would be great if we have a standard repo id that "just works" when you use it.

[mm65x]

done

[mm65x]

uploaded the converted model here: https://huggingface.co/mlx-community/ACE-Step-v1.5-turbo-mlx

repo id is mlx-community/ACE-Step-v1.5-turbo-mlx if you want to point docs/examples at a standard hosted MLX repo.

[Blaizzy]

Please name it by dtype or quant type

mlx-community/ACE-Step-v1.5-turbo-bf16

Or

...-4bit

The convert script already adds tags that identify the model as MLX 👌🏽

[mm65x]

done

[lucasnewman]

🚀