Towards Understanding the Cognitive Habits of Large Reasoning Models