feature request: add support for minicpm4 model (a ultra fast model in small size)

[Remember2015]

as tech report mentioned, it is really fast.
mlx version already existed: https://huggingface.co/mlx-community/MiniCPM4-8B-4bit