On the Completeness of Invariant Geometric Deep Learning Models