Convert method for only for pytorch/safetensors to MLX

Hi,

Currently the convert method implementation converts from pytorch/safetensors to MLX format:

CONVERT

We would like the convert to also do the reverse e.g., convert from MLX to pytorch/safetensors.

mlx_lm.convert --help

usage: mlx_lm.convert [-h] [--hf-path HF_PATH] [--mlx-path MLX_PATH] [-q] [--q-group-size Q_GROUP_SIZE] [--q-bits Q_BITS]
                      [--quant-predicate {mixed_2_6,mixed_3_4,mixed_3_6,mixed_4_6}] [--dtype {float16,bfloat16,float32}] [--upload-repo UPLOAD_REPO]
                      [-d]

Convert Hugging Face model to MLX format

options:
  -h, --help            show this help message and exit
  --hf-path HF_PATH     Path to the Hugging Face model.
  --mlx-path MLX_PATH   Path to save the MLX model.
  -q, --quantize        Generate a quantized model.
  --q-group-size Q_GROUP_SIZE
                        Group size for quantization.
  --q-bits Q_BITS       Bits per weight for quantization.
  --quant-predicate {mixed_2_6,mixed_3_4,mixed_3_6,mixed_4_6}
                        Mixed-bit quantization recipe.
  --dtype {float16,bfloat16,float32}
                        Type to save the non-quantized parameters. Defaults to config.json's `torch_dtype` or the current model weights dtype.
  --upload-repo UPLOAD_REPO
                        The Hugging Face repo to upload the model to.
  -d, --dequantize      Dequantize a quantized model.

We can add a flag --convert-to where we specify to which format we want the conversion to happen.

Regards,

Start with llms/mlx_lm/convert.py and run mlx_lm.convert --help to understand the existing Hugging Face-to-MLX options. Define how a --convert-to option should select MLX-to-PyTorch/safetensors conversion, including the expected output path and supported cases; done means the reverse conversion works and the command help documents it.