Demystifying Structural Disparity in Graph Neural Networks: Can One Size Fit All?