* This way an even tighter space bound can be found for float writes.

* The server buffer creation does not involve the clients,
  hence it is better to call MPI_Win_create early.

* The approach with cpp relies on GNU tools, i.e. is not portable.

* In a follow-up step it will also be used for source lists.

* random is subject to arbitrary disturbances because libraries like to
  assume it can be used indiscriminately.

* The structure is more generally useful than only server-side
  destination index lists.

* This will need further changes later on, once more parts of CDI use
  size_t instead of int.

* ppm_extents.h is pulled in by ppm_uniform_partition.h but since the code
  contains variables of type PPM_extent, it is more consistent to include
  it explicitly.

* Also fixes too late initialization of pointer.

* Also intercept nc__enddef for parallel access.

* With the reductions being pre-computed, this is only necessary when
  writing to a NetCDF stream that could only serially be opened.

* This way the distributed grid can return the corresponding values
  (first and last value in X- resp. Yvals and increment later on
  without communicating with ranks holding other shards of the grid.