ReFACT: Updating Text-to-Image Models by Editing the Text Encoder