Multimodal Diffusion Language Models for Thinking-Aware Editing and Generation