llamapilot is very slow

[madebyollin]

loading the model takes ten thousand years. we should do it once, instead of many times.

[madebyollin]

ggml-org/llama.cpp#61 seems like progress (interactive mode now allows multiple queries to the same process). but it looks like interactive mode persists the prompt over time, so might still need to modify llama.cpp to allow issuing multiple independent queries to the same process.