Reasoning and Sampling-Augmented MCQ Difficulty Prediction via LLMs