In MethodDNN one could start training with a batch size larger than the
number of events in training leading to non-sensical output. TMVA now
properly warns (fatal warning) and suggests a fix (decrease batch size).