Convex Formulations for Training Two-Layer ReLU Neural Networks