How to estimate RPS requests to LLM based on message frequency from users, session duration, and number of users, and compare it with the model's throughput capacity?