Optimize categorical crossentropy gradient update #91
Description
The gradient and hessian update of the categorical crossentropy loss computes p_k k times, but it only needs to compute it once (see scikit-learn/scikit-learn@9e68984 which led to serious improvement in fit time).