Scaling Spatial Intelligence with Multimodal Foundation Models