Setup cross-validation pipeline

[mnick]

We start to have enough data to explore model selection via cross-validation. Ideally, we'd want a pipeline where we can sweep over hyperparameters with submitit (using gridsearch or nevergrad) and try to forecast cases for a held-out set of the last 2-3 days. We'd then retrain the the model using these hyperparameters on the entire dataset to produce the final forecasts.