Cache Monarch matrix + causal mask for faster inference

Upload model.jl with huggingface_hub