The State of Sparsity in Deep Neural Networks