Removed several nullify routines. Also made sure that several subroutines that copy derived data types that do not have a finalizer call the corresponding *_end routine explicitly and have the corresponding argument is intent(inout).

Remove the INCR macro in order to stick as much as possible with Fortran, so that the code is not too much obfuscated.

The packing for sending and unpacking from receiving is done with a
kernel that implements a general reordering according to a map.

The implementation again supports both normal MPI and CUDA-aware MPI.

Replace real(8) by FLOAT when possible.
Remove some single precision routines, as this is not supported anymore

This reverts commit 3a970b43.

With this approach, the dot products with offsets can be effectively
overlapped. The infrastructure of using different streams could also be
expanded to other parts of the code.

In this case, the Nvidia MPS daemon has to be used. Otherwise, the
different processes cannot access the GPU memory of the other processes
as needed for the CUDA-aware MPI implementation because the memory
belongs to different contexts. When using MPS, the different contexts
are merged under the hood.

In any case it is probably a good idea to run MPS.

Added packed, unpacked and device packed versions of density_calc_accumulate for non-collinear spin.

This is needed to ensure the cublas functions are executed in the same
stream as the rest of the CUDA kernels and thus are synchronous to the
other operations.

These functions get the device pointers from C and put this pointer into
a fortran pointer that canbe used like a regular pointer or array for
MPI calls when CUDA-aware MPI is enabled.

The configure-time switch --enable-cudampi and the variable
CudaAwareMPI are introduced to enable or disable the use of CUDA-aware
MPI.

parse_is_defined and messages_obsolete.

* Pass the parser object down the call tree.

The derived data types are now declared with all components private by
default with public components indicated explicitly, or a comment is added
 when the components are public by default in which case the private components
are indicated explicitly.