feat(marshaling): also marshal during predict

[sebffischer]

Summary:

• setting unmarshal = FALSE not ensures that all models are in unmarshaled form (resample(), benchmark()). This is just easier to reason about
• models are now also marshaled for prediction, i.e. for callr encapsulation and parallel prediction (I made a wrong assumption + somehow did not properly test this assumption)
• the marshaling of models for prediction happens automatically during Learner$predict(), but the model will be restored to its state before prediction.
• the logic for the marshaling in the worker was refactored and is now much easier to understand
• The LearnerClassifDebug now got an argument called check_pid that is FALSE by default. If this is set to TRUE, the unmarshal_model() method for model_classif_debug sets the marshal_pid slot of the model. The private $.predict() method of LearnerClassifDebug then checks that the marhsal_pid is equal to Sys.getpid(). This lets us test that the unmarshaling is executed correctly for the prediction.