Proto-Value Networks: Scaling Representation Learning with Auxiliary Tasks