Representation Purification for End-to-End Speech Translation