Semantic Abstraction: Open-World 3D Scene Understanding from 2D Vision-Language Models