This is one more step towards using the parallelized calculation of the
stepfunction in the inp.xml code path.

This is the first step to enable the parallelization for parts of this code
as it is provided by Uliana.

Esepecially jij%nmagn was already needed in rw_noco which is called in
a subroutine of dimens. Therefor it has to be initialized before dimens
is called. I hope there are no side effects of this relocation.

a) fixed CuDOSXML test
b) made compile with PGI
c) change testscript to work without "do"

- Relaxation of atom positions now works as it should.
- The code also becomes clearer.

- if inp or inp.xml file is present take it and be happy
- if both are present, give an error unless -xml option is given
- if none is present stop (before opening files)

The feature is used by invoking fleur with -hdf_cdn -dd a-b
to delete all densities with index between a and b. Using
-dd a only deletes density a.

Note: The required space for the densities will not yet be given
back to the system. The size of the file stays the same. A
solution to this problem may be the usage of the tool h5repack.
It should be available everywhere.

...this is very incomplete at the moment.

Some call in main/fleur_init.F90 was at the wrong place. Fix by Daniel.

The command line argument is something like "-sd 5" (to select density no. 5.)

Now the attribute minDistance is read in from the input file into
the variable input%minDistance. By default this variable is 0.0 but
it can be set to finite values in the inp.xml file. It is supposed
to be an additional optional convergence criterion to stop calculations
when they are converged.

removed: weight, natd, nkpts,nkptd

- made l_zref a local variable no longer in type sym
- changed the setup of the parallelization (mpi/mpimakegroups is obsolete)
- added new command line switches yet to be documented :-)