Stare at What You See: Masked Image Modeling without Reconstruction