Remove texture reference support in the CUDA

Remove texture reference support in the CUDA

Only required for legacy CC 2.0 / Fermi hardware we drop support to
simplify code and improve maintainability. The impact on the Fermi
support will only be performance (as direct memory fetch will be used).
Additionally, all builds will use multiple complication units now,
single compilation unit support has been kept for now and will be
removed later (if it simplifies code).

Follow-up expected/made possible:
- remove single compilation unit mode for CC 2.0;
- remove barrier that protects from race cleanup in
  pme-load-balancing.cpp;
- merge regular and Fermi NB kernels.

Change-Id: Ic3677b7beaff1b4b9dc927d955940f7e779c2a41