XGBClassifier conversion to onnx gives wrong predictions if dataset has only 2 categories (xgboost 2.0)

[veselina-marinova-cresset]

Following the procedure for onnx conversion here: https://onnx.ai/sklearn-onnx/auto_tutorial/plot_gexternal_xgboost.html , I found that the probabilities when loaded from the onnx file are wrong for a dataset with less than 3 classes. The issue is present for xgboost 2.0 but not for xgboost 1.7