Add tests for backward compatibility

[mllg]

We should store snapshots of learners trained with older versions of mlr3 and predict with them in the unit tests to ensure backward compatibility.

[sebffischer]

I think we should write an FAQ entry how to circumvent these issues in case they appear.
Maybe we could include a link to this FAQ entry in warning / error message that appears when there is a version conflict.
What do you think?

[mllg]

Storing a trained learner and task is completely reasonable and we should ensure that they are also compatible with future versions of mlr3.

There are also cases where a version mismatch happens accidentally (multiple hosts during parallelization), but I don't see what fix we would write down in a FAQ which we cannot perform automatically if everything is versioned.