Mitigating Degree Bias in Graph Representation Learning with Learnable Structural Augmentation and Structural Self-Attention