Adding RateLimiter to LLM unit tests, plus improving system-chat prompt to work better for both llama 70b, gpt-3.5-turbo, and gpt-4o-mini.