Using machine learning potentials to advance zeolite property prediction beyond the accuracy of classical force fields