Add better dataset with stresses (or virials)

[frostedoyster]

As found out in #296, we only have 10 alchemical structures to perform tests on strain gradients.
However, these are very few (because they're large, so we kept 10 so as to limit their size), and this breaks our models since they complain about missing atomic types in the train dataset when compared to validation and test.

We should find a dataset with small periodic structures with stresses or virials, and add the missing test in #296