prefix caching in llama.cpp with MLX

[dockridge]

I'd approach it in a boring order. If you're on vLLM, I'd confirm the version first and then reduce variables until the behavior is obvious.