This used to be in the base class as default implementations, but I think
it is nicer to have it in a derived class.  In particular, this will
eventually allow to do the FD approximation with high-precision number types.

[[Imported from SVN: r9842]]

It sometimes creates problems with ADOL-C.

[[Imported from SVN: r9841]]

[[Imported from SVN: r9840]]

Gauss-Seidel converges equally good/bad as IPOpt, but it is much faster.

[[Imported from SVN: r9839]]

[[Imported from SVN: r9838]]

Amazingly, this really does have a measurable effect on the overall computation
speed.  Assembly times for the global Hessian and gradient for the Cosserat shell
energy problem drop by about 5% (!)

[[Imported from SVN: r9837]]

That's obviously more efficient when solving one linear equation
with multiple right hand sides.

[[Imported from SVN: r9836]]

Previously the energy norm was used, but as we are solving non-convex
problems the Hessians may be indefinite, and then the EnergyNorm is not
a norm.

[[Imported from SVN: r9835]]

[[Imported from SVN: r9834]]

[[Imported from SVN: r9833]]

Previously we called hess_mat (vector-mode second derivatives) with the
canonical basis vectors, and then rotated the resulting matrix.
This is wasteful, as more directions are used than necessary.
This patch uses the minimal number of directions.

[[Imported from SVN: r9818]]

[[Imported from SVN: r9817]]

The hasObstacle_ and obstacles_ fields are now single arrays.
The corresponding fields for the entire grid hierarchy is assembled
within dune-solvers now.

[[Imported from SVN: r9816]]

This almost halves the assembly time for the Hesse matrix.  Nice!
Nevertheless I still keep the scalar-mode code around, hidden
behind and #ifdef.  Higher-order vector-mode in ADOL-C appears
to be not quite as reliable as the rest, so I want to be able to
easily compare results with the scalar code.

Also, unlike the scalar code, the vector-mode one again computes
the full Hessian in the surrounding Euclidean space.  Only afterwards,
that full Hessian is multiplied by the tangent space basis vectors
to obtain a Hesse matrix on the tangent space.  To obtain better
efficiency one could have ADOL-C compute the Hessian in the tangent
directions directly.  However, ADOL-C sometimes segfaults... when
doing that.

[[Imported from SVN: r9815]]

[[Imported from SVN: r9814]]

[[Imported from SVN: r9813]]

Newer versions of ADOL-C do not define 'overwrite' anymore, hence there
is no risk of macro clashing.

[[Imported from SVN: r9798]]

[[Imported from SVN: r9786]]

[[Imported from SVN: r9785]]

Instead, only let ADOL-C evaluate the Hessian in the directions
of the vectors of the orthonormal frames.  The result is the same,
but we get away with fewer calls to ADOL-C, which, in my Cosserat
shell example, shaves off about 10% of the assembly time.

[[Imported from SVN: r9779]]

[[Imported from SVN: r9777]]

Otherwise, all processes will write into the same file --> not a good idea.

[[Imported from SVN: r9776]]

[[Imported from SVN: r9771]]

Previously, the matrix was always assembled using a P1-basis, even if we used
it to measure errors of vectors pertaining to P2 or other spaces.  Nobody noticed
this, because the h1SemiNorm routine would not fail if the vector was larger
than the matrix.   Sigh...

[[Imported from SVN: r9770]]

That makes for slightly simpler code.

[[Imported from SVN: r9768]]

[[Imported from SVN: r9767]]

For a general grid, such loops are only possible for codimensions 0
and the grid dimension.

[[Imported from SVN: r9766]]

[[Imported from SVN: r9765]]

[[Imported from SVN: r9764]]

[[Imported from SVN: r9763]]

[[Imported from SVN: r9762]]

GlobalP2Mapper is intended to be for P2 finite element spaces, what
GlobalIndex is for P1 spaces: a numbering for all degrees of freedom
that is globally unique across all processors.

[[Imported from SVN: r9761]]

[[Imported from SVN: r9760]]

Iterators over entities are only mandatory for vertices and elements.
But here the user may also request indices for other entities.
Hence one cannot use a straight iterator.

[[Imported from SVN: r9759]]

No idea why valgrind didn't complain about this.

[[Imported from SVN: r9758]]

Even if there is only one processor.  Because even then the local
and global vectors differ in their ordering.

[[Imported from SVN: r9757]]

Currently only for quadrilaterals.  I'll implement triangles as soon
as I need them.

Unfortunately, VTK appears to only support 8-node quadrilaterals
(without an element dof).  So the element dofs get ignored currently.

[[Imported from SVN: r9756]]

With this patch we start to differentiate according the the function
space bases of the input data.  We write P1 data as is, and downsample
P2 and P3 data onto a P1 space.

[[Imported from SVN: r9755]]

[[Imported from SVN: r9754]]

This makes visualizing wrinkles much easier.

[[Imported from SVN: r9753]]