Training-Free Structured Diffusion Guidance for Compositional Text-to-Image Synthesis