Interface FuseOptions

mlx_lm.fuse --dequantize: write dense weights for every quantized module (fused or not) and drop the config's quantization block.