Deductive Verification for Chain-of-Thought Reasoning in LLMs