The multigrid hierarchy is built in two different ways, depending on whether
we use a P1NodalBasis for the discretization, or anything else.  Previously,
this switch was made using the preprocessor.  The patches changes the code
to use std::is_same instead.

[[Imported from SVN: r10046]]

This is much cleaner than handing over the Grid only, and then trying
to guess the basis using the preprocessor.  After all, the
RiemannianTrustRegionSolver calls the assembler, so it may know the
function space basis.

[[Imported from SVN: r10035]]

[[Imported from SVN: r10033]]

[[Imported from SVN: r9995]]

[[Imported from SVN: r9994]]

There is no need to have our own implementation for this, we had it purely for
historical reasons.

Also, this change fixes a subtle bug:  Our own implementation used AlienElementLocalBasis,
which returns zero if the LocalBasis is evaluated outside of the reference element.
However, this happened sometimes due to rounding error on grids with very small quadrilateral
elements.

[[Imported from SVN: r9991]]

[[Imported from SVN: r9990]]

[[Imported from SVN: r9965]]

Previously, the number of multigrid transfer operators was always equal to
the number of grid levels minus one.  That worked out correctly when doing
first-order FE spaces.  However, for higher-order spaces, the up-most transfer
operator transfers from the higher-order space to a first-order space
_on the same grid level_.  But since the total number of transfer operators
was grid levels minus one, this implied that the coarsest grid leve 
was never used, and the coarse grid solver would operate on the second-coarsest
grid.

This patch adds the extra transfer operator in this case.  That means we now
call the coarse grid solver on smaller problems, which makes a noticeable
run-time difference.

[[Imported from SVN: r9964]]

[[Imported from SVN: r9963]]

Normally, the two different level grid views.  Previously we used the same one
to compute the localToGlobal renumbering for boths rows and columns.  Of course
that screws up the numbering.

I am not quite sure when that got introduced.  I may have gotten in fairly recently,
after the upstreaming of the GlobalIndexSet classs.

[[Imported from SVN: r9962]]

[[Imported from SVN: r9954]]

Because it is not an index, it really is a mapper.

[[Imported from SVN: r9953]]

[[Imported from SVN: r9936]]

In particular this means having to carry mappers for the local indices around.
This makes the code more verbose, but it also makes it easier to see what is
happening.  While coding this stuff I keep having doubts about whether all
the renumbering works correctly...

[[Imported from SVN: r9925]]

The local copy was to get the thing developed and working.  Now that it is
working we upstream it into dune-grid.

[[Imported from SVN: r9922]]

The IndexSet interface class from dune-grid doesn't have it, and we want
to be like the interface method as much as possible.

[[Imported from SVN: r9921]]

[[Imported from SVN: r9918]]

[[Imported from SVN: r9917]]

[[Imported from SVN: r9909]]

[[Imported from SVN: r9890]]

dune-solvers actually does provide a way to set these matrices directly.
I simply didn't find it until recently.

[[Imported from SVN: r9885]]

[[Imported from SVN: r9881]]

[[Imported from SVN: r9878]]

[[Imported from SVN: r9868]]

[[Imported from SVN: r9854]]

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

[[Imported from SVN: r9839]]

[[Imported from SVN: r9838]]

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: 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]]

[[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: r9765]]

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

[[Imported from SVN: r9757]]

The Dirichlet nodes are handled directly via the ignoreNodes_ field.
I don't know how the matrix modification code ever got in here.

Removing the code doesn't seem to make any difference.  I find that
a bit surprising: I would expect multigrid and IPOpt to work better
if the matrix is truly symmetric.

[[Imported from SVN: r9742]]

Much easier to use than the previous API.

[[Imported from SVN: r9735]]

[[Imported from SVN: r9734]]

[[Imported from SVN: r9732]]