Cache Monarch matrices + causal mask for faster inference

Upload checkpoint.jl with huggingface_hub