Augment with Care: Enhancing Graph Contrastive Learning with Selective Spectrum Perturbation