* Also rename corresponding C sources.

* This is only one step mandated by the brain-dead behaviour of the
  NAG and PGI compilers respectively.

- Removed unnecessary, misleading comments
- initialized boundary of s

power_method now returns with 0 if starting vector is 0.
chebyshev iteration returns with right solution if lambda_max = 0 or lambda_max = lambda_min

- IBM xlf doesn't allow for this kind of functions
- The matrix stencil now has a shift property to simulate the old behavior
- Additional shifted versions of all matrix stencil were necessary

It was forgotten to pass the "global" variable to the arr_norm_2() function

- all solvers now use the functions from linear_algebra like arr_dotproduct arr_norm_2 etc.
- in linear_algebra.f90 renamed mat_* functions to arr_* functions (more meaningful)
- all spectral methods now take an additional extent as argument

The basic idea of the halo variants was to save communication when a larger
halo is used and the system is solved globally (meaning no Schwarz).
Tests showed that this didn't result in great advantages because in the case of CG
when communication is necessary two halos need to be transfered instead of one.
Although block-jacobi based preconditioners make obviously problems in this setup.
A workaround for this, (decreasing the range of the preconditioner block, so that they
don't overlap) was never tried because it most definitly will result in more iterations.
The punch line to all of this is that it just didn't pan out.

- lambda_maxiter and _tol for functions related to eigenvalue calculation
- cheby_miniter and _checkrate for break condition in chebyshev iteration

- added an apply_stencil function that does what it says
- renamed get_lambda_min_max to calc_lambda_min_max
- new get_lambda_min_max returns the values from the config

- added new file solver_public.f90

CG with different preconditioners tested

This was necessary after rebase

Chebyshev iterations works now without preconditioner
testing for chebyshev with preconditioner is pending

Thereby the behaviour of get_lambda_max was slightly changed. From now on
it doesn't change the input vector x any more.
In order to get the corresponding eigenvector to a eigenvalue use power_method
instead.

To tidy up solver.f90 a bit, global_sum, global_norm_2, calc_rel_res, calc_abs_res
and global_dotproduct were moved to linear_algebra.f90

Thereby global versions of dotproduct and norm were implemented

Internal datastructures are now in solver_internal.f90 which shouldn't be imported
by external applications.