Philippe 04/04/2017: OpenACC: porting to OpenPOWER + optimisations

* use data zones instead of data create (to prevent crashes on OpenPOWER
  with PGI up to 17.3)
* use atomic and force loop parallelisation for COUNTJVxD_DEVICE subroutines
* force parallelisation of some loops that were not automatically parallelised
* added NOKRR7 define key to remove optional argument part of the code
  of OpenACC kernels (to prevent crashes on OpenPOWER)