logits is a (batch, classes) tensor of raw scores and target is a 1-D tensor of integer class indices. Return the mean cross-entropy loss: convert each row of logits to a probability distribution, then average the negative log-probability assigned to the true class across the batch.