Keras CNN models

First - this package is amazing and I am so relieved to see someone has been working to make importance diagrams work for Keras. This is an underdeveloped area and extremely gratifying to see.

[In the Predictor.R script, lines 106 to 110](https://github.com/christophM/iml/commit/409838af43cd9303dc6a119697360c84863ea5ab), only data frames can be imported into the Predictor$new function. This precludes the use of list arrays, which are needed for more complex models such as CNNs (specifically one dimensional CNN's, where you just expand dims to 3 to make it work). Example code is as follows:

` listarrays::expand_dims(x_train_proto, 3)
`
This means that importance diagrams can't be made for CNN or LSTM models in R with this package - I'm curious if there are any ideas of a workaround? One thing I can think of is for the FeatureImp$new function to detect a CNN model in Keras and wrap appropriately. The other would be to import a proper data frame, but also a listarray just for the purpose of getting the Keras model going, and linking any column name functions to it. Happy to help with this, but not sure where to start.