Specialized CUDA kernels lose a reference to the cache

[gmarkall]

Quick notes on this based on a verbal description from @guilhermeleobas: The specialize() function doesn't create the specialized dispatcher with any reference to caching, so caching is broken for specialized dispatchers. (See line 702 of numba/cuda/dispatcher.py).