Reduced-Order Neural Operators: Learning Lagrangian Dynamics on Highly Sparse Graphs