Code 11.56 and 11.57 issue with negative log probabilities from softmax

For the code in 11.56 and 11.57, I had been trying to replicate the results using the `pm.math.softmax` method, however, I was getting errors during sampling due to it producing negative values.

Adjusting `p_ = at.nnet.softmax(s)` to be `p = pm.math.softmax(s, axis=1)` resolved the problem.