Efficient Adaptation of Visual Foundation Models for Wood Defect Segmentation via Instance-Linking and Feature Disentanglement