Rethinking Message Passing Neural Networks with Diffusion Distance-guided Stress Majorization