Self-adaptive vision-language model for 3D segmentation of pulmonary artery and vein