Weakly Supervised 3D Open-vocabulary Segmentation