Geometry-Informed Neural Operator for Large-Scale 3D PDEs