The Unreliability of Explanations in Few-shot Prompting for Textual Reasoning