feat: neuralnet classif learner from old mlr2

[AnnaNzrv]

closes #479

This PR implements binary classification only, following the approach from the previous version of mlr2.

The following steps have been completed to ensure proper functionality:

• The added learner(s) are sufficiently tested
• All the CI tests are passing (including the CodeFactor)
• Update docs with devtools::document()
• Updated the NEWS.md field to include the addition of the learner
• Not modify anything not related to the new learner
• Contributor in the DESCRIPTION of the R package

[AnnaNzrv]

While extending the learner to support multiclass classification should generally be possible, as outlined in an example of the neuralnet documentation, there is some unexpected behavior. Specifically, when applying neuralnet::neuralnet to tasks like the wine classification (tsk("wine")), the model predicts the same response for all observations. This issue requires further investigation if we decide to extend the learner to multiclass classification.

Example:

task = tsk("wine")
pars = list(act.fct = "logistic", linear.output = FALSE, err.fct = "ce")
target = task$target_names
class_names = task$class_names
data = task$data()

# Create formula based on neuralnet documentation
lhs = paste0("(", eval(target), "== \"", class_names, "\")", collapse = " + ")
formula_str = paste(lhs, "~ .")
formula = as.formula(formula_str)

model = invoke(
      neuralnet::neuralnet,
      formula = formula,
      data = data,
      .args = pars
    )

newdata = task$data()  # ordered_features(task, self)
invoke(predict, model, newdata = newdata)
#> Prediction matrix with only duplicated rows