Detect NaNs early

[cvanaret]

See https://github.com/cvanaret/Uno/issues/432#issuecomment-3544285116

The strategy I'm going for: Each subproblem solver should have two copies of its EvaluationSpace: one at the current iterate and one at the trial iterate. In the trial_evaluation_space, we evaluate the objective gradient and Jacobian at the trial iterate. We then check if some values are NaNs. If so, the trial iterate is rejected. If not, we accept it and swap the current and trial evaluation spaces, so that we reuse the evaluations as current evaluations.