prefix caching in llama.cpp: worth it?

[bytesupply76]

That matches what I was seeing. The pattern around tail latency has been pretty similar for me. The biggest change for me was simplifying the stack instead of adding one more optimization layer.