Extend neural network classes TATi can create

[FrederikHeber]

TATi is so far limited to (fully) feed-forward networks. However, for image classification Convoluted NNs are better, for signal processing Long-Short-Term Memory (LSTM) NNs.

Tensorflow 1.9 offers tf.keras as the main module for easy prototyping, see http://tensorflow.org/versions/r1.9/programmers_guide/keras.

• can we use the keras module to replace (parts of) our 'neuralnetwork.py' with tests still passing?
• If yes: allow the user to specify python scripts using tf.keras to setup network
• If yes: prepare scripts for several archetypical networks (FFNNs, CNNs, LSTMNNs) that the user can refer to by string name (and some options).