Reorganize how recommendation engines are contstructed.

[rat-nick]

I have ran into an idea that a recommender engine is a composition of the model and the input adapter.

The model

It's an ML model that should only be concerned with performing a forward pass on data.
For KNN should be a vector in feature space, for VAE should be a vector of implicit user ratings etc.

The adapter

The adapter is responsible for bridging the gap between inner and outer IDs, converting item ratings into feature space vectors etc.

This way we can have a variety of different engines built upon a KNN model like content KNN with cosine similarity as the distance, CF KNN with cosine and jaccard distance and so on.

So my idea is that the Engine should be a composition of a Model object and an Adapter object.