Efficient Manifold-Constrained Neural ODE for High-Dimensional Datasets