Learning Generalizable Features for Tibial Plateau Fracture Segmentation Using Masked Autoencoder and Limited Annotations