vLLM adds MiniMax M3 and DiffusionGemma support

Day-zero support for 1M-token multimodal reasoning model and first diffusion LLM in production-grade inference serving.