A Geometric Insight into Equivariant Message Passing Neural Networks on Riemannian Manifolds