Softmax Cross-Entropy & Its Gradient

Implement cross_entropy(logits, targets) returning (loss, dlogits):

• logits: shape (n, C)
• targets: integer class indices, shape (n,)
• loss: the mean negative log-likelihood (a Python float)
• dlogits: gradient of the mean loss w.r.t. logits, shape (n, C)

The beautiful result: $\frac{\partial L}{\partial z} = \frac{1}{n}(\text{softmax}(z) - \text{onehot}(y))$.

Use log-sum-exp for numerical stability.