Try to reduce the number of datasets that require manual download

> Currently, 41 canonical datasets require manual download. I checked their scripts and I'm pretty sure this number can be reduced to ≈ 30 by not relying on bash scripts to download data, hosting data directly on the Hub when the license permits, etc. Then, we will mostly be left with datasets with restricted access, which we can ignore

from https://github.com/huggingface/datasets-server/issues/12#issuecomment-1026920432