Both Semantics and Reconstruction Matter: Making Representation Encoders Ready for Text-to-Image Generation and Editing