Optimizing Reasoning Efficiency through Prompt Difficulty Prediction