Implementation of a fallback.learner option

[benesom]

A rather general implementation to be used with learner$new can look something like this
learner_fallback(original, fallback) { … }
original, fallback are both learner objects

1. estimation fails for original learner - > use fallback learner
2. prediction fails for original learner -> use fallback learner (therefore, we always need to estimate the fallback learner)

2 checks

1. convergence
2. error during estimation -> tryCatch

We should review if we need this fallback behaviour only for the superlearner. I can only come up with rare use-cases (mostly for debugging or experimentation) where one would like this feature to be available for a single learner to be used as the outcome.model in cate. This can obviously be done by using learner_sl with a single base learner and fallback learner.

[benesom]

Note that the prediction method will not fail for learner_ constructor models once we implemented #184. Should the fallback.learner then replace NAs in the predictions of individual learners before the predictions are ensembled?