Allow TNNMG to use Solvers::LinearSolver object

See merge request !5

Clang complains that this must be initializes from a constexpr
(which is maybe not the case here). However, there's no need for
constexpr here, so let's calm down clang and make the test pass again.

While I'd like the tested interface more, it's not implemented
yet leading to a failing test

Vector is derived from the functional

In dune-solvers!13, the dune-solvers variant of IsNumber was removed.

Feature/new tnnmg simplex solver

See merge request !3

Not just the lower limit.

This file was introduced originally, because the AcceleratedGSStep
class would only accept linear solvers for the linear correction,
but no iteration steps.  The FixedStepLinearSolver could be used
as a shim to practically turn an iteration step into a solver.

Recently, the TNNMGStep class got generalized and now accepts
LinearIterationStep objects directly, in addition to linear solvers.
This makes the FixedStepLinearSolver class obsolete.

If ever somebody else should need it, it should be in dune-solvers
anyway.

There are two alternatives now for the linear correction:
You can either provide a solver or an iteration step.

Now, the step can be anything that inherits from Solvers::IterationStep.

This was a non-standard extension introduced without discussion.

As a consequence of this removal, the FixedStepLinearSolver class
loses its first template argument, too.

Because, in the dune-solvers world, you currently need to access
it directly to call 'setProblem' on it.

That's not particularly pretty, but it's the way it is.  A general
improvement of dune-solvers here is welcome.

This includes it into the regular dynamic-polymorphism class
tree of dune-solvers.

In e01898cc the template paramter V was removed. This commit updates the
AcceleratedNonlinearGSStep accordingly.

Take it from the Functional type instead.

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.

* 'subDiff' is now called 'subDifferential'
* It returns its result in the return value, not in its second argument
* The method 'domain' also returns its result in the return value

And get rid of the file newsumfunctional.hh, which was never intended to
be a permanent solution.  With this patch, the file sumfunctional.hh
contains both the new and the old implementation, and you can switch
between the two with USE_OLD_TNNMG.

In the new tnnmg implementation, the calling method is expected to know
whether a functional is quadratic or not.  In case of a quadratic
functional, the caller will not call bisection anyway, but it will
rather call a dedicated solver for quadratic problems.

This patch also introduces the preprocessor variable USE_OLD_TNNMG.
When this variable is defined, the old tnnmg implementation is
reenabled. This will simplify running old and new codes side by side
for a while, which I find helpful for debugging.

Previously, if the subdifferential at the initial iterate would contain zero,
no message would be written at all.  This made it more difficult to follow
the programm output, because it was impossible to see how many times
the bisection method was called.

It is not actually used.