Denoising Diffusion Semantic Segmentation with Mask Prior Modeling