because otherwise the cannons algorithm is not usually tested

also adding wrapper for the backwords step

the output matrix has to be set to zero before calling the routine!

whether cannon algorithm can be used. If not, switch back to the
original implementation

this commit does not work, the generalized EVP problems fail from 2
reasons:
- Cannons algorithm does not work for the processor grid shape we use at
  the moment in tests (It only works if num. of process columns is a
  multiple of num. of process rows, which is not true in ELPA for 2 mpi
  tasks)
- Cannons algorithm does not work as integrated to ELPA for larger
  matrices (not clear why at the moment)

There are 2 new tests, which should work
- test_cannon.c: it tests the new algorithm without going through
  Fortran, as it has been delivered
- test_c_bindings: it tests transfering a 2D fortran matrix to C and
  back, as it is done with the cannons algorithm in the normal ELPA
  tests

Rewrote the perl-script for the dependency generation to also accept
input file names from stdin, in order to avoid excessively large
argument lists.

so that it works if one of bandred / band_to_full steps are done on GPU
and the other one on CPU