Add support for non-linear Zoo Models

[braun-steven]

Is your feature request related to a problem? Please describe.
ZooModels from Deeplearning4j such as UNet, NASNet etc. have a graph-based, non-linear architecture.
Since the Dl4jMlpClassifier holds all layers in a linear/sequential manner, it is not capable of parsing these models.

Idea
Create a Dl4jZooModel classifier that mimics the Dl4jMlpClassifier behavior and can load arbitrary zoo models but does not allow to change the model architecture, except for the output and input.