Closes: #67

Importing the `vtk` python module has significant overhead.  Using
only one python invocation (and thus only importing the module once)
reduces this overhead significantly.

Add freestanding function referenceElement(geometry)

As suggested in #65, this merge request adds a freestanding function
`referenceElement(geometry)` that returns a reference element for a grid
geometry **by value**. This will work for all grids with geometries
based on the standard reference elements. For grids that want to return
something else, the behavior can be changed by providing an overload for
`referenceElement(const Geometry&, const GeometryImplementation&)` in an
appropriate namespace.

Moreover, it is also possible to get the type of the reference element
for `Geometry` using `Dune::ReferenceElement<Geometry>`.

Both the change to value semantics and
`Dune::ReferenceElement<Geometry>` rely on dune-geometry!52. I've
also tried to adapt all of dune-grid to the new interfaces (in
particular, the value semantics and the deprecation of
`Dune::ReferenceElement<ctype,dim>`) - but no guarantees.

I had to build in a pretty ugly hack to make the thing work on Debian 8,
which is contained in a single commit and should be reverted once we
drop support for GCC 4.9.

This merge request closes #65.

See merge request !194

Test for idset injectivity

See merge request !191

When UGGrid is configured with MPI, then the ids produced by
dune-uggrid are injective across all codimensions and
GeometryType.  This is the way it should be in Dune (modulo
the 'copy' rule).  However, when MPI is not used, then two
entities of different codimension can have the same id.
To fix this, we now encode the entity codimension into the
highest two bits of the id.  This was already done in the
code previously, but not for all codimensions, and not in
a very legible way.

This patch implements the 'id' method for entities of codimension 1
in a three-dimensional UGGrid.  The implementation is incomplete:
it will work in general, but it will return wrong results if a facet
has a copy on a coarser level.

Nevertheless, the current implementation makes the test pass again.
I will improve the implementation (and sharpen the test) in a
subsequent patch.

The computation of UGGrid ids is a bit complicated, and needs
some cleanup.  As a first step, this patch moves the cpp #ifs
that distinguish between MPI and non-MPI builds into the
ugwrapper.hh file.  That way each #if happens only once in
the code.

remove support for old UG library

Closes #66

See merge request !199

Make 'size' return the same type as 'IndexSet::size'

See merge request !196

Explicitly deducing size_type from the IndexSet
(as in MCMGMapper) fails for some 'incomplete-type'
problem.  But 'auto' is still almost as good:
only the exported size_type is missing, but few
people are expected to need that.

[Cleanup] Remove mpivtktest.in.

See merge request !198

Previously, the SCSGMapper class had 'int' hard-coded as the
return type of the 'size' method.  This method changes it
to whatever the underlying IndexSet is using for its 'size'
method.  I think this should settle the infamous
'signed-vs.-unsigned' discussion at least for this method.

Additionally, the new type is exported as 'size_type'.

This file was needed with the autotools buildsystem to run vtkest with two
ranks.  The autotools buildsystem is gone, and with cmake the same effect is
achieved from CMakeLists.txt alone.

The libraries are already added in the `foreach` loop on l. 45-51.

It is already used in `dune_register_package_flags` and the compiler
definitions should be the same there and in `dune_add_flags`.

Closes: #66

Bugfix in dgfparser when using tetgen.

See merge request !197

is fixed by passing the absolute value.

Previously, the MCMGMapper class had 'int' hard-coded as the
return type of the 'size' method.  This method changes it
to whatever the underlying IndexSet is using for its 'size'
method.  I think this should settle the infamous
'signed-vs.-unsigned' discussion at least for this method.

Additionally, the new type is exported as 'size_type'.

GCC 4.9 on Debian 8 trips over the `friend auto referenceElement(...)`
in Geometry. I can't really debug that as I don't have Debian and my
Homebrew GCC 4.9 compiles the code just fine. So this is a brute-force
workaround that sidesteps the problem explicitely for GCC < 5 (I hope we can
abandon that antiquity sooner rather than later anyway).

The hack is rather ugly, as I couldn't find a way to make
`referenceElement()` a friend of Geometry (the compilers tripped about
all kinds of other problems when adding a trailing return type), so we
introduce a helper struct that takes over friend duties.

This patch replaces (hopefully) all occurences of the following
deprecated usage patterns:

- capturing and passing around reference elements by reference
- assuming that any reference element returns data by reference
- usage of the deprecated `Dune::ReferenceElement<ctype,dim>`

It also switches to using the new freestanding `referenceElement()`
function in most places.

This patch adds a function `referenceElement(Geometry)` that returns
the reference element for a grid geometry. In order to allow grid
implementors to return custom reference elements (e.g. for polyhedra),
it performs a second-level lookup by calling `referenceElement()` with
both the geometry facade and the geometry implementation stored inside
the facade class.

The default implementation of the second-level `referenceElement()`
function just looks up the default reference element for the
GeometryType of the geometry. Grid implementations that want different
behavior should overload this function for their specific geometry
implementation(s).

The YaspGrid implementations uses our custom overloads for printing
std containers, but never included them, because they were indirectly
included through the reference elements, which in turn included
`<dune/common/array.hh>`. As that include switched to `<array>`, stuff
broke in weird ways. So include what you use!

IndexSet, IdSet: make default constructor protected

Closes #64

See merge request !192

Closes: #64

According to the class documentation of the IdSet class, ids must
be different for any two different entities of a grid (of any level
or codimension), except when one entity is a copy of the other.

This test adds a unit test for this feature.

.gitlab-ci.yml: update environments

See merge request !190

- Added:   Debian 10 with gcc in C++17 mode
- Removed: Debian 8 + Backports with clang

Reference: http://lists.dune-project.org/pipermail/dune-devel/2017-July/002224.html

add accessor for underlying `VTKWriter`

See merge request !186

add missing include for dverb

See merge request !185