With the recent changes to bigunsignedint, the index cannot
be initialized with a negative value anymore. We change this
to 1 to overcome the problem.

This commit together with the recent bugfixes by Ansgar
fixes the AMG code on master.

Since 0c697d8d messages of size zero
are no longer sent. This can reduce the number of requests we have to
wait for.

The storage of bigunsignedint was changed to std::uint16_t in commit
541b3dfc. This implements the same
change for the MPITraits<bigunsignedint<k>>

Note that this assumes that
  MPITraits<std::uint16_t>
is defined. This is the case as long as uint16_t maps to one of the
standard integer types (unsigned char/short/int/long).

Thanks to Markus for pointing this out.

Reference: https://dune-project.org/flyspray/index.php?do=details&task_id=1657#comment6393

Before this commit communcation would be initiated even on empty
interfaces. While this does not lead to overwriting data or producing
segmentation faults, it does lead to accessing the first component of
empty vectors to get a pointer to pass to MPI functions. This produces
error output with valgrind and other memory debuggers. Therefore with this
commit we will first check whether the interface contains values and do
nothing if it doesn't. We also augmented the test to catch the case wher only
some processes have an empty interface.

This check was added to improve the error message returned when using
MPI without correctly initializing it (e.g. by creating a parallel
grid), see FS#1612 for further information.

@Markus: too much CMake?! ;-)

There is not need to send message that contain no data. It might even
produce problems on  some implementation. Therfore this patch always
checks the size and simply sets the request to inactive (aka
MPI_REQUEST_NULL) if no data would be send or received.

The patch is to be cherry-picked to master if Christian approves.

Like we do for Autotools.
This is only supported for CMake 3.0 or newer. Oder versions ignore
the property and report the test as failed.

"Makefile.in", "*.o", etc. need only be listed in the toplevel .gitignore, the
rules are applied recursively.

Test programs etc. should be listed as "/program" in their directory's
.gitignore, so they are not accidentially ignored in a lower level directory.

This very old gcc bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=10200 is fixed in clang but not in icc or gcc, see FS#1505.

In order to simplify fallback handling between code with and without MPI
support, this patch makes the specialization of CollectiveCommunication
for MPI communicators default-constructible by defaulting the underlying
MPI communicator to MPI_COMM_WORLD, which should be reasonable for most
scenarios.

As the non-specialized version for sequential code is already
default-constructible, this allows grid constructors to accept a
CollectiveCommunication object regardless of whether MPI support is
enabled or not and to provide a default value for this parameter.

This patch resolves FS#1497.

CollectiveCommunication serves as a fallback implementation, when
code that is written for parallel computations is executed in sequential.
The template parameters BinaryFunctions are no-op then, but they are
still there to define the interface. However, in a generic (parallel)
implementation these will be given. Including binaryfunctions.hh here
avoids error messages due to no-op template parameters whose names
could not be found by the compiler.