Make dataset loaders on-the-fly

[cthoyt]

I think it would be better to have the dataset download and processing happen client-side, then use pystow to store the results in a reliable place. This would also allow the TWOSIDES and DrugBank datasets, which require random negative sampling, to be used with multiple random seeds, e.g. to investigate the robustness of results. Further, it would allow for a more idiomatic dataset loader that's extensible to new datasets

Depends on:

• Add the DrugComb and DrugCombDB data cleaning #50
• Clean up DrugBank and TWOSIDES importers #57
• Provide base class for dataset loaders #59