Add an assertion to the CNN example

[jpivarski]

Sometimes, this can happen:

I don't think we need to impose a fixed random seed (as we do in the autoencoder example). We can let the results vary from day to day, but only publicly display successful fits, as we do with Minuit (see 1bf2f2a). For the CNN, the assertion can be on the "diagonalness" of the confusion matrix, maybe a minimum trace. (The total scale is fixed by the number of entries in the sample.)