Weakly-Supervised 3D Spatial Reasoning for Text-Based Visual Question Answering