Can Language Models Perform Robust Reasoning in Chain-of-thought Prompting with Noisy Rationales?