Learning physics-constrained subgrid-scale closures in the small-data regime for stable and accurate LES