REFER: An End-to-end Rationale Extraction Framework for Explanation Regularization