integrate with other llms (e.g., llama, gemini)

[sunil-2000]

as each turn is defined only by a prompt, we should be able to plug in different llms and use puzzles as interesting benchmark for comparison between them.