Think How to Think: Mitigating Overthinking with Autonomous Difficulty Cognition in Large Reasoning Models