Improve implementation of `jtprod`

Following #379 , the implementation of `jtprod` [L. 475](https://github.com/JuliaSmoothOptimizers/NLPModels.jl/blob/289b81ceea0961cb60a6e536d32311114b09405b/src/nlp/api.jl#L475) allocate by default when the problem has linear and nonlinear constraints.

The optimal way is to re-implement this method when creating a `NLPModel`, but we could think of alternatives.
For instance, having a 5-arguments `mul!` with `jtprod`.