[WIP] Move tests to new JuMP nonlinear interface

[frapac]

We have one remaining issue: the MathOptNLSModel depends on the legacy nonlinear interface to formulate quadratic residuals which are not supported by default.

We should add this support explicitly if we do not want to break NLPModelsJuMP. This would require some non-trivial changes in MathOptNLSModel though.

[blegat]

I have a plan to add NLS support with the new API, see 94a1327
The plan is to update MOI.Nonlinear to allow arrays (as done currently in ArrayDiff by copy-pasted much of MOI.Nonlinear and modifying it). Then, NLPModelsJuMP can just detect the expression sum((...).^2) and build an NLS model.
So my idea is that if the model fits, it creates a QuadraticModel (see #230), otherwise, if it detects that the objective is a LS, it creates an NLS model and otherwise it creates a generic NLP model.