Think or Not? Exploring Thinking Efficiency in Large Reasoning Models via an Information-Theoretic Lens