Handle nonnumeric predict_proba class labels

[FlorianPfaff]

Summary

• Avoid converting model.classes_ through the active PyRecEst numeric backend when selecting the positive predict_proba column.
• Preserve the numeric positive-label behavior for classes_ = [1, 0].
• Fall back to column 1 for nonnumeric class labels instead of risking backend conversion errors.

Testing

• Not run locally: the execution container cannot resolve github.com, so I used the GitHub connector for repository writes.
• Added a regression test for nonnumeric classes_ fallback in tests/test_association_models.py.