It is basically a combination of AcceleratedNonlinearGSStep and
TNNMGAcceleration, but I think it makes the code easier to understand,
even though it does not get shorter.

In addition, the new TNNMGStep class takes nonlinear smoother in form
of a shared_ptr to a dune-solvers IterationStep again.  I don't see
why we need static polymorphism here.