Statistics for Defending Graph Convolutional Networks against Dynamic Graph Perturbations via Bayesian Self-Supervision