Beyond the Laplacian: Interpolated Spectral Augmentation for Graph Neural Networks