Cross-Entropy with Label Smoothing

Training on hard one-hot labels pushes models to be over-confident. Label smoothing spreads a little probability to the other classes: with smoothing $\epsilon$ and $C$ classes, the target for the true class is $1 - \epsilon + \epsilon/C$, and every other class gets $\epsilon/C$.

Implement smoothed_targets(labels, n_classes, eps) and smoothed_cross_entropy(logits, labels, eps=0.1): the mean over the batch of $-\sum_c t_c \log p_c$, using a numerically stable log-softmax. Also return the gradient with respect to the logits, which is (softmax - targets) / N: (loss, grad).