MIRAGE: Evaluating and Explaining Inductive Reasoning Process in Language Models