Less is More: Towards Efficient Few-shot 3D Semantic Segmentation via Training-free Networks