Diversified Sampling Improves Scaling LLM inference