Contrastive Decoding Improves Reasoning in Large Language Models