`MIN` mode via acquisition function

[AdrianSosic]

This PR changes the behavior of NumericalTarget in MIN mode in that the minimization is lo longer implement by negating the computational representation of the target but placing an objective on the acquisition function.

This approach avoids the problem that otherwise the surrogate would be trained on inverted targets, resulting in inverted predictions, which hinders us from exposing the surrogate as-is to the user. Also, it cleanly separates the objective (i.e. "minimization") from the target representation.