Support quantization mode

[awni]

The latest MLX adds a quantization mode to quantization operations. In mlx-lm we add the mode to the config so we know what type of quantization it is. See e.g. https://huggingface.co/mlx-community/InternVL3_5-1B-4bit/blob/main/config.json#L173

That flag causes models to break with our examples here.

[davidkoski]

Perfect, ml-explore/mlx-swift#273 should add the back end support we need.