Factor common infrastructure between CUDA simulator and hardware target implementations

[gmarkall]

As the simulator evolves along with the CUDA target, there is an increase in logic / coda duplication between the two - e.g. as suggested in https://github.com/numba/numba/pull/4975/files#r364319727, and particularly around the operation of real and fake device arrays.

It would be good to factor this shared and duplicated logic into a common, or shared module, from which both the real target and the simulator are composed.