Scene-Driven Multimodal Knowledge Graph Construction for Embodied AI