Remove hardcoded warp_size == 32 assumption from PME GPU

Remove hardcoded warp_size == 32 assumption from PME GPU

The PME OpenCL host code now queries the warp_size aka max
execution width from the device. The CUDA/unit test handling
of constants/defines is adjusted. With this, only the PME
device-side code should be relying on warp_size define
(which will be passed to the compiler in the OpenCL
implementation).

Change-Id: Ic2d1512e04d36861590b13e02b5bd7a87240f9e2