Cross-Entropy with an Ignore Index

Padded sequence batches use a special target (often -100) for positions that shouldn't count. Implement cross_entropy(logits, targets, ignore_index=-100):

• logits has shape (N, C), targets shape (N,)
• the loss is the mean negative log-likelihood over non-ignored positions only (stable log-softmax)
• if every position is ignored, return 0.0

Also implement cross_entropy_grad(...) with the same arguments, returning the gradient with respect to logits (zero rows for ignored positions).