Improve transformation tests

[tyarkoni]

The transformation tests need some work. Specifically:

• Instead of using the bundled ds005 dataset, which is large, we should start using fake datasets we can initialize more quickly and control more precisely. See test_dropna for an example.

• Per discussion in Rename transform broken #318, the coverage of the transformation tests (in analysis/tests/test_transformations.py) is lacking: most of the tests only test using a dense variable or a sparse variable, and not both. It would be good to test both everywhere they're applicable.