CriticBench: Benchmarking LLMs for Critique-Correct Reasoning