Active Self-Training for Weakly Supervised 3D Scene Semantic Segmentation