KV cache disk offload

[ha-seungwon]

I'm trying to run llama.cpp on a small machine, but the KV cache is too large. Instead of pre-allocating the KV cache in memory as a buffer, is there a way to offload it to disk?