Log-Softmax and Negative Log-Likelihood

Cross-entropy is usually computed as log-softmax followed by NLL, which avoids taking the log of a tiny probability.

Implement:

• log_softmax(z) over the last axis: $z - \log\sum e^{z}$ computed with the log-sum-exp trick
• nll(log_probs, targets): mean of $-\log p_{i, t_i}$ for integer class targets