problem to anticipate here is that not all resamplings create a prediction for every input row (#216 will get relevant here), and some may create multiple predictios (does mean averaging usually make sense?)

This is rather urgent though, as the current version blocks e.g. mlr3spatiotempcv resampling and makes the PO non-extensible which popped up in at least one project.
We could perhaps allow all mlr_resamplings and just document that this might break?

In general, another possibility would be to

• in case of e.g. holdout, just fill up the rest of the predictions with ǸA`
• in case multiple predictions exist just create additional columns as needed.
  the current return format is a data.table anyway.

filling with NA is probably a good idea. I don't like multiple columns because the number of output columns must be the same in train() and predict. I guess if it's supposed to be quick we can just mean(). It would be nice though to do something sensible with se and prob. (But I guess we haven't solved that problem for PipeOpRegrAvg / PipeOpClassifAvg either.)

The number of output cols: We can just store the number of outputs in train and enforce the same length in test.
Aggregating using mean or m̀ode` might also work (perhaps this could be an option),