Combining Wasserstein-1 and Wasserstein-2 proximals: robust manifold learning via well-posed generative flows