kernel selection. This logic will be further complicated by allowing
individual routines to be run on GPU

Previously, matrix a has been kept on the device at the end of bandred
and then copied to host later in redist_band (called from tridiag_band).
However, it seems to be more convenient to do it at the end of bandred
(since it has to be done anyway). Band to tridi is thus now not using a
on the device at all.  However, we keep it in the interface, sice it
might be usefull in the future.

elpa2_trans_ev_band_to_full_template

NOT testing with gpu,
NOT testing with no-assumed-size

tile size is used in elpa1_tridiag and elpa2_bandred.
Iits value is computed based on block size and least common multiple of
number of processes rows and colums, but then can be enlarged to some
multiple of this value. How much it should be enlarged can now be
influenced by min_tile_size parameter.

avoiding calling both check eigenvalues AND check eigenvalues and
eigenvectors for analytic eigenvector test

if provided a a fourth parameter, all correctness checks will be
skipped. Can be usefull for repeated performance tests.

it seems, that the new compiler interprets order parameter in reshape
command in the opposite way to its predecessor.