Create true end-to-end integration test designed to catch insidious pipeline bugs

One of the (many) goals of this project is to prevent data science user errors, like temporal leakage or bad metric calculation. We should take care to avoid reproducing such errors here, or using constituent components (ie collate) in a way that can create those errors.

To help with this, we should create an end-to-end test with a synthetic dataset and grid/scoring config containing some landmines that could trip up a badly-designed version of this pipeline. Creating this dataset is not trivial, but should be worth it