feature: let people add training rows multiple times for the same row_id, but for different model_ids

[ClimbsRocks]

basically, you should be able to train a model 10 times.

at serving time, we should be able to just grab the rows by their row_id (ignoring model_id)

at training time, we should be able to have feature values stored for each iteration of the model we train, to make sure each one conforms to what we expect.

this will definitely require some custom logic. but i think it'll be helpful.