cuda.cooperative: Eliminate link= argument to @cuda.jit decorator when it becomes possible

[shwina]

Currently, cuda.cooperative relies on passing any additional code we wish to link with the user-defined kernel via the link= argument to @cuda.jit:

warp_exclusive_sum = cudax.warp.exclusive_sum(numba.int32)
@cuda.jit(link=warp_exclusive_sum.files) # <<<<<<<<---- awkward 
def kernel(data):
  data[cuda.threadIdx.x] = warp_exclusive_sum(data[cuda.threadIdx.x])

Ideally, the user wouldn't have to do this. There's work underway within Numba CUDA that would avoid this, instead letting us declare device code beforehand, and Numba would link it whenever it's used (example). When that functionality is available, we should switch to using it.