OptimalThinkingBench: Evaluating Over and Underthinking in LLMs