Add reminder about validation curves' dependency on generalization performance metric

In the [Overfit-generalization-underfit notebook](https://inria.github.io/scikit-learn-mooc/python_scripts/cross_validation_validation_curve.html) we present the validation curve of a `DecisionTreeRegressor()` using the Mean absolute error to score the model.
This could be a good opportunity to remind people that scores and errors cover different ranges of values and, therefore, training and testing curves can swap their relative positions depending on the evaluation method. I think this could partially help clarifying the notion of good vs. bad fit.