A difference of two pointers was casted to unsigned which can lead to
overflow on 64-bit systems.

previous implementation was not thread/stream safe, since it used constant
memory

new implementation doesn't use any global objects, so it is thread/stream safe

Add GPU models in seperate folder, only old models are supported and they were replaced to make the CPU interface work.
This solves the undesired consequence of breaking the GPU interface.

move hog::set_up_constants from constructor to compute method

if user changed CUDA device between constructor and computation,
some variables were uninitialized

The architecture has been correctly determined earlier in this
file. No need to do it again.