Understanding Inductive Biases of Graph Neural Networks