[Bug] issue with max on GPU

[jbcaillau]

@ocots on the rocket / goddard case which has a max, the (mayer) cost is actually minimised when using the combo exa + GPU; which explains the wrong result. same for other pb with max on the bench, i guess. looking further

cc @AmielMetier @HediChennoufi

[jbcaillau]

@AmielMetier @HediChennoufi

• we first need to deal with [Dev] Change max to min - for all problems OptimalControlProblems.jl#178
• then re-run the bench on CPU + GPU with tol=1e-8 (1e-6 is not enough to get neat controls)

[jbcaillau]

now solved in main branch of ExaModels MadNLP/MadNLP.jl#518 should be available in the the next release