Training MLPs on Graphs without Supervision