Respect `MOI.AutomaticDifferentiationBackend()` in NLP

[klamike]

The current code has SparseReverseMode hard-coded: https://github.com/jump-dev/DiffOpt.jl/blob/bf2c4c389ecd6ae63356e292b74a632b0517fa68/src/NonLinearProgram/nlp_utilities.jl#L94-L109

Perhaps backend should be set based on MOI.AutomaticDifferentiationBackend()? Besides the SymbolicMode, it would also allow to use DiffOpt with custom evaluators.

[joaquimg]

We do need an attribute for setting that.

[blegat]

We could use MOI.AutomaticDifferentiationBackend by default. We could have a different attribute in case the user wants a different AD than the solver but using the same as the solver to start with would be a good first step