[FEA] Create an operator that aggregates previous feature values while grouping

[karlhigley]

Is your feature request related to a problem? Please describe. I wish NVTabular could help me roll up the last ten or so item ids that a user interacted with.

Describe the solution you'd like Create an operator based on groupby("field").shift(N) functionality from cudf and pandas.

Describe alternatives you've considered Another way to do this would be to build a full list of all interacted items and then slice it down to the items that were interacted with previously somehow (based on either timestamps or the position of the current item in the list.) This seems pretty inelegant and inefficient though.

Additional context This would be useful for enriching the user features available for use in the query tower of a retrieval model, particularly since many publicly available datasets don't contain much in the way of user information.