qnvars, update_obj! and trustregion

[pkofod]

It should be possible to only evaluate the objective to check the trustregion. It should also be possible to only update on success if quasi-newton. For Newton's method this might not be a good idea because the hessian can often be evaluated with the gradient and objective cheaply. Hope for many accepts. We should accept and return QNVars, not individual things.