Rewriting softmax calculations fixes two rounding errors that distort training

Barron and colleagues report that precision-preserving formulas remove spurious gradients in high-confidence classification and closely matched teacher-student networks.