On the Use of Unrealistic Predictions in Hundreds of Papers Evaluating Graph Representations