BLIP-Diffusion: Pre-trained Subject Representation for Controllable Text-to-Image Generation and Editing