Theano optimization as special case for Classifier

[alexjc]

When using theano.nnet's implementation of Lasagne's categorical_crossentropy, it uses a different codepath for float32 probabilities instead of 1-of-n int32 onehot encoding. The second codepath has a GPU optimization that makes it about 50% faster on GTX970.

First experiments in making this work failed though... needs a bit more investigation.