Large Diffusion Vision-Language-Action Model with Unified Multi-Modal Instruction and Generation