Error Bounds for Deep Learning-based Uncertainty Propagation in SDEs