Add recommend-and-measure API

[mdekstrand]

Right now, we only support measuring top-N recommendation lists that have been pre-generated. For unlimited-length lists on large data files, this is infeasible.

We need a 'recommend-and-measure' API that computes metrics over recommendation lists without needing to store the lists in memory. There are two ways we could do this:

• add a method to RecListAnalysis
• add a batch.recommend_and_measure function

[mdekstrand]

When we implement #503, measurement can be an alternate collector.