[TASK] Extract unrolling logic out of Aggregation into Unroll & make it mask-aware

[marcromeyn]

In a session-based recsys setting, you often have contextual-features (1-D features like time-of-day, device-type, etc.) & session features (2-D features, a feature per timestep of the sequence). We typically merge these features by unrolling the 1-D features into the shape of the session-features:

This logic is now part of the aggregations (here to be specific). We would like to break this out into a new class (working name Unroll).

One reason for this is that we want to make use of the standard way in Keras to propagate masks (more info can be found here). This new class should implemtent compute_mask in order to alter the masks (it should add the mask info for all the contextual features).