Understanding Non-linearity in Graph Neural Networks from the Bayesian-Inference Perspective