`GraphLearner` should implement custom `$configure()` and `Graph` should also implement `$configure()`

[sebffischer]

mlr3's $configure() allows to set both hyperparameters and fields.
It would be nice if GraphLearner provided its own $configure() method that also allows to set fields of PipeOps, something like

graph = po("pca") %>>% lrn("classif.rpart")
glrn = as_learner(graph)

glrn$configure(
  classif.rpart.predict_type = "prob"
) 

Currently this does not work, because glrn does neither have a field nor parameter called classif.rpart.predict_type.

Analogously, it would be nice if the same was possible for Graphs, which currently don't have a $configure() method.

[mb706]

We could set up $configure so that it tries to call $configure() in contained objects whenever it does not find a field or hyperparameter in the class itself. So GraphLearner calls Graph's $configure() (to be implemented), Graph$configure() strips off a PipeOp id prefix and calls the corresponding PipeOp, PipeOpLearner delegates to the Learner etc

[mb706]

From our internal discussion: The predict_type example actually already works, since GraphLearner's predict_type is already delegated through the active binding:

graph = po("pca") %>>% lrn("classif.rpart")
glrn = as_learner(graph)

glrn$configure(predict_type = "prob") 
glrn$graph$pipeops$classif.rpart$learner
#> <LearnerClassifRpart:classif.rpart>: Classification Tree
#> * Model: -
#> * Parameters: xval=0
#> * Packages: mlr3, rpart
#> * Predict Types:  response, [prob]
#> * Feature Types: logical, integer, numeric, factor, ordered
#> * Properties: importance, missings, multiclass, selected_features,
#>   twoclass, weights

but there may be other fields for which this does not work.