The Role of $n$-gram Smoothing in the Age of Neural Networks