multiple LLM clients

[dovinmu]

This generalizes to allow different clients that can do LLM inference.

• removes AIPL's dependence on llama.cpp
• new assumption is that everything will by default be over HTTP
• adds the idea of a client (similar to HELM) which handles the inference
• adds clients for OpenAI, GooseAI, and a custom HTTP server that is compatible with the openai library

[anjakefala]

This is so much better and clearer, Rowan! Tysm for refactoring!

Did you test on a local model? I tested with OpenAI, and once I removed the requirement for the environment variable it worked great!

[dovinmu]

@anjakefala yep, I was using the testing code at the bottom of clients.py to make sure each of those worked (though didn't test permutations of env, thanks for the catch!).