[DNN Training] Add API/support for Model Interpretability of DNNs

Once we have the foundational DNN training API stabilized, we should add DNN Model Interpretability APIs such as methods of type Saliency Methods and Feature Attribution (FA).

For further details, this is a high level summary:
https://towardsdatascience.com/interpretability-of-deep-learning-models-9f52e54d72ab