Controllable Text-to-Image Synthesis for Multi-Modality MR Images