Automated generation and surrogate modeling of 3D urban wind simulation based on physics-informed graph neural networks: a two-stage framework