On the Stability of Graph Convolutional Neural Networks: A Probabilistic Perspective