Is Chain-of-Thought Reasoning of LLMs a Mirage?