Scaling MLPs: A Tale of Inductive Bias