[ImageClassification Transfer Learning] Support additional DNN architectures

Just a reminder that we'll need to support more powerful architectures and specialized architectures in order to cover most customer scenarios.

For instance, NASNet (notably nasnet_large and pnasnet_large), which can provide extra precision. Or for lighter models targeting slower processors, MobileNet V1 or V2 architectures, or nasnet_mobile.

See:
https://www.tensorflow.org/hub/tutorials/image_retraining#other_model_architectures

I wonder if we should try to make a generic catalog based on "needs" or "image types" instead of architecture names that most .NET developers won't be familiar with such as the previous names or Inceptionv3, ResNet, etc. Those names don't mean anything to most .NET developers.

OPEN PRE-TRAINED MODEL:
The other feature is to be able for the user to provide their own model, but for that, we'd probably need an API with many more details, such as tensor names?