Preconditioning Newton–Krylov methods in nonconvex large scale optimization