Add a buffer in NLPModelMeta for jtprod!

[amontoison]

@tmigot
I just remarked that jtprod! allocates if nlin > 0 and nnln > 0.
I think we should allocate a buffer of size nvar in NLPModelMeta to store an additional Jtv (linear or nonlinear) and then do the sum.
-> https://github.com/JuliaSmoothOptimizers/NLPModels.jl/blob/main/src/nlp/api.jl#L503-L523

cc @frapac

[tmigot]

Yes, it is a known issue. Our approach so far has been to recommend to re-implement jtprod on top of jtprod_lin and jtprod_nln. As in most of the applications, it is possible to do it, but that's very application specific.

Adding a buffer in NLPModelMeta is not without consequences as it will increase the memory usage.

Overall, I would be in favor of documenting this further.

[tmigot]

I close this as a duplicate of #384