Parameterize LLMClient for people to use whatever LLM they want

[kamath]

We only run evals on 4o, o1, and Sonnet rn, and we don't want to unnecessarily complicate evals. However, we don't want to stop users from experimenting with different LLMs. Thus, we should expose LLMClient and allow people to pass in an LLMClient of their choosing if they want to use providers like Ollama, Groq, Gemini, etc. Relates to #350, #183, #252