Use L-BFGS if the Hessian of the Lagrangian is not available

[amontoison]

With the last release of NLPModels.jl (0.21.6), we can specify what is implemented or not in the API.
We should use L-BFGS by default if nlp.meta.hess_available is false.
I already added this kind of behavior in NLPModelsIpopt.jl (JuliaSmoothOptimizers/NLPModelsIpopt.jl#148).