metatensor/metatensor#770 added the ability for models to request a non-strict NL from the MD engine, with the explicit goal of improving pair_style metatensor by removing the need to update the NL we give to the model at each step. We should implement the corresponding functionality in this repo.