This PR puts Netflix prize data in a format where it can be conveniently queried and subsetted. Call me crazy, but I tend to think in JOINs, and I appreciate having an RDBMS sweat the memory management details for cases where not everything will conveniently fit in core. So sqlalchemy is managing some sqlite tables. They are safe to DROP, or alternatively just rm out/movies.sqlite and they're all gone and will be rebuilt on the next run.

This PR also trains a LightFM model on (some of) the prize data, and shows how to make predictions. They are, frankly, unimpressive, but I figured we need something in the code base that demonstrates how to do it, so we can all build upon it. The tests run quickly, to support an interactive edit-run cycle.

Change etl("mv_00*.txt", max_rows=1_000_000) from 1 M rows to 101 M rows to ingest everything. And then it will be preserved -- we skip ETL if a table is already populated.

The "give me two movies you like and I will recommend some more" is frankly still aspirational at this point, but we're not far from it. That's where two_movies_rec_test.py got its name.

There are some design notes in make_recommendation.py, which may serve as the basis for future PRs. Once this merges down to main, it's possible we will view the lightfm_recommendation.py module as obsolete and slated for deletion. It did offer me some inspiration for the current work.

I'm happy to accept any and all comments, but please be sensitive to what is in-scope for the current PR as opposed to a subsequent PR. It would be useful to get this merged down, with light edits, this week.