CoMPaSS: Enhancing Spatial Understanding in Text-to-Image Diffusion Models