* collSpec is meant to be a dumb data collection, none of its member should
  be subject to change.

* gcc 6.4 does not realize posix_memalign initializes ptr on some
  Linux systems.

* The previous implementation had the destructor also release the mapping.

* That was problematic, because the internals of objects depended on
  the external mapping unnecessarily.

* This also fixes a potential double-free issue discovered by Jan
  Frederik Engels.

* Not yet created resource handles are by definition unused.

* This applies to NetCDF where the first write contains invariants that
  are not present in later writes.

* Especially create the latitudes symmetrically from pole to pole.

* Bull X MPI incorrectly creates the shared file pointer in many cases.
* This fix alleviates the problem but cannot fully restore sane behaviour.

* Versions 4.1 to 4.3.2 of netcdf do not provide programmatically useful
  pre-processor predicate if the pnetcdf layer is available and report
  success when the nc_create call actually failed.

* For statically decomposed streams, the xmaps can be precomputed under the
  assumption that all elements are written.

* 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.