[FEA] Make `cuda.core.Buffer` recognized by numba-cuda

[leofang]

[brandon-b-miller]

@leofang just getting my head around the requirements here. Is the goal to make it so that one can pass cuda.core.Buffer objects as arguments of numba kernels? Or are you envisioning a world where one can write x = cuda.core.Buffer(...) in a kernel and expect back a heap allocated buffer for the cuda thread to use for something downstream in the kernel? Or something else?

If it's the former I think all we might need to do is support __cuda_array_interface__ for cuda.core.Buffer.