Print out the kernel name and also some more detail about the dimensions
that are problematic.

This does not impact performance, but helps in debugging

Co-authored-by: Cristian Le <cristian.le@mpsd.mpg.de>

This is recommended by the CUDA best practices guide, see
https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#thread-and-block-heuristics
Make sure that the limit of threads per block for the corresponding
kernel is respected (though it is unlikely to be smaller than 256).

causing an error. This is now properly captured by the code.

Replace `kind_oct_m` with `intrinsic :: iso_fortran_env`. Access to the fortran intrinsics is also available through `global_oct_m`.

Replace all instances of _i4, _i8, r4 and _r8. Note, this affects routine names as well as precision suffixes.

Update OPTION variables to also use standard fortran precision, _int64

Replace declarations using kinds aliases with fortran standard variables. Note, formatting has not been preserved, but it's only cosmetic.

... messages_print_with_emphasis

The use a complex as two doubles required to have the state index running first
in the GPU kernel.

fix indentations
remove trailing white spaces

The routine was not working for integers, causing the code to stop is one tries to initialize all GPU buffers to zero.

Also show the host name. This allows easier identification of ranks and
devices to hosts. Also get rid of the leading zeros in this output.

Batchifying the calculation of the commutator of the position operator and the DFT+U(+V) term. This allows to compute the current fully batchified.

Avoid potential integer overflows when computing the product of
pack_size along its dimensions or with other, potentially large numbers.
Add some more low-level routines to allow calling with i4 and i8
integers.

For large batch sizes, the multiplication of the number of grid points
with the number of states in a batch could overflow. Change the
integers involved to 8-byte.