The Convex Relaxation Barrier, Revisited: Tightened Single-Neuron Relaxations for Neural Network Verification