Doe jacobian remake

[Osburg]

This PR adds:

• Computation of objective function gradient using autodiff and formulaic's symbolic differentiation of model matrices. Now supports arbitrary model terms and not only polynomial model terms of order <= 3. Significant speedup of gradient evaluation in case of large models with a large number of experiments.
• Restructuring the objective function and its gradient. A new class DOptimality now contains both constraint evaluation and constraint jacobian evaluation.
• __call__() and jacobian() methods for NChooseKConstraint using narrow Gaussians as discussed in DOE: NChooseK constraints #145 .
• Option in doe to pass NChooseK constraints to IPOPT (using the new constraint evaluation and jacobian) instead fixing a corresponding number of variables in each experiment to zero before the optimization. Possibly to be removed in the future since the new option does not seem to work nicely so far.

[jduerholt]

Out of curiosity: What do you mean with the new nchoosek option does not seem to work nicely? What is happening if you use it?

[Osburg]

About the nchoosek constraints not working properly: I explained my findings here.

[jduerholt]

Looks good to me. Thanks!

[jduerholt]

@dlinzner-bcs @DavidWalz : Do you also want to have a look, or should I just merge?

[jduerholt]

I think, we should just merge it ;)