On the Limitations of Model Stealing with Uncertainty Quantification Models