Previously, this test used a situation with geodesic interpolation
between two points on the sphere that are almost antipodal.
This made the TargetSpaceRiemannianTRSolver crash, because the
weighted sum of squared distances can have indefinite second
derivatives in such a situation.  In principle, the
TargetSpaceRiemannianTRSolver should handle this, but nowadays
it is not actually a TR solver anymore, but a plain local
Newton method.  Maybe I should revert back to trust-region.

I added a comment to targetspacetrsolver.cc that explains the
situation a little.

There is really no point in having this method.  Before it is called,
localgeodesicfefunctiontest already checks whether the derivative
is correct (by comparing with an FD approximation).  If a derivative
passes this test it MUST be tangential.  No need to check that
explicitly.

The computation of the A-orthogonal matrices was wrong,
and for some reason nobody ever really noticed this.
One reason could be that in GFE applications A is frequently
very close to the identity, and then you wouldn't notice
the bug that much.

Anyway, localgeodesicfefunctiontest.cc passes now.

It is not clear what this method was supposed to test,
because the CosseratEnergyStiffness class does not
actually implement a the gradient.

To prevent confusion with a class of the same name in
dune-localfunctions.

In favor of the new ones in dune-common.

All of a sudden it gets confused with a class of the same name
in dune-localfunctions.

For some reason I must have forgotten to check in this file when
first converting to the cmake build system.

Because it really is only an energy.

Because it is not a stiffness, it is really just an energy.

It is still part of the official matrix interface, but I doubt
that anybody actually uses it.  I'd rather see blocklevel gone
and replaced by modern C++ introspection.  Let's remove it here
and see who complains.

The dune-common 2.7 release finally adds all those standard
binary addition, multiplication, etc operators for FieldVector
and FieldMatrix.  Hence we do not have to have our own
implementation anymore.  We still keep it for backward-
compatibility with dune-common 2.6, but guard it behind
cpp conditionals.

Previously, they all inherited from LocalGeodesicFEStiffness,
which was weird, because none of them implemented first or
second derivatives as you would expect from something called
'stiffness'.

It's not clear what it is supposed to be testing at all.

And make it actually test something.

Previously the code only tested whether the AverageDistanceAssembler
computed gradient vector and Hesse matrix.  Now the test actually
tests whether these are correct.

This makes the methods AverageDistanceAssembler::assembleHessian
compile again.  And it should be a wee bit faster two, because
matrix symmetry is now exploited better.

Because a Stiffness does provide a first derivative.

LocalFirstOrderModel will be an abstract base class for models that
implement an energy and a first (not no second) derivative.

A LocalGeodesicFEStiffness 'is a' LocalFirstOrderModel, which 'is a'
LocalEnergy.

... rather than a LocalGeodesicFEStiffness.  That is a much
cleaner design -- A 'stiffness' produces first and second
derivatives for an 'energy'.

Currently, all implementations of variational problems inherit
from LocalGeodesicFEStiffness, which requires gradient and Hessian
implementations.  However, most of these implementations really
only implement the energy, and leave computations of first and
second derivatives to algorithmic differentiation.  To make
this clearer, this patch introduces an abstract base class
LocalEnergy for all implementations that really only have
an energy.

Throw out the AmiraMesh support instead.  I have not seen
anybody using Amira for years.

It combined computing rod configurations and their discretization
errors.  Nowadays the way to do that is to use rod3d.cc for the
rod configurations, and compute-disc-errors.cc for the errors.
Hence rod-eoc.cc is not needed anymore and can go.