[Dev] Test enzyme backend for ADNLPModels

[jbcaillau]

@PierreMartinon follow up to control-toolbox/OptimalControl.jl#481 (comment)

• test solve with backend=:enzyme (does forward over reverse but uses Enzyme instead of ReverseDiff for reverse)
• solves the issue documented above, and should be faster on other cases: to be tested
• if conclusive, use if as the new default instead of backend=:optimized
• currently fix only available in branch main of ADNLPModels

[PierreMartinon]

Ok, I'll look into it !
NB. Following our discussion, probably not a new default, unless on older Julia versions (LTS ?)

bump with a related question: would the enzyme backend be more suited to handle something like the variable step solvers from OrdinaryDiffEqs ?
(cf JuliaSmoothOptimizers/ADNLPModels.jl#280)