REF/Review interface to optimizers

[josef-pkt]

one refactoring:
allow that methods used as objective functions, score and hessian are optional
e.g. GLM has gmmobjective and gmmobjective_cu
M estimators like RLM don't have a loglike.

Review:
base.model.Model.fit has
xopt, retvals, optim_settings = optimizer._fit(f, score, start_params, ..
how much does this differ now from minimize directly? (our wrapper precedes scipy minimize)
It looks like the main part of our work is after this call when checking Hessian.

One difference to scipy is in our own optimizer methods, e.g. newton and L1 for discrete.

(related: We still have Wald inference In LikelihoodModel instead of a more generic class)

[josef-pkt]

GLM case: it would be useful to allow for deviance as objective function
score and hessian should be the same in this case as with loglike.

It would make it easier for tweedie, e.g. bug in elastic net because we don't have proper loglike #7476

Adding an option for selecting the objective function in the optimizer code when creating the lambda functions would, I guess, not be difficult to implement.

The could be an instance attribute, if we want the pick objective function only during fit (GLM), or could be a class attribute if everything is based on the same objective function (GMM).