Better Zero-Shot Reasoning with Self-Adaptive Prompting