Improving 3D Few-Shot Segmentation with Inference-Time Pseudo-Labeling