Open minimal datasets without downloading the full one

[juntyr]

In the online lab example, https://gist.github.com/juntyr/ea85adab5df19141a84939f2d322697f, we have to download the full dataset to open the minimal one. This barely works (we have just enough memory for it) but that's not a sustainable solution. We should bring back a way to download the small datasets directly, at least for the datasets that support it.

[treigerm]

Follow-up on this, we should be able to host the pre-processed tiny datasets on Zenodo. This should make it maximally reproducible.