Reasoning Robustness of LLMs to Adversarial Typographical Errors