[BUG] RMMNumbaManager broken when Numba is using NV CUDA bindings

[brandon-b-miller]

Describe the bug
Numba can switch back and forth between use of its own ctypes bindings for cuda objects, and the new cuda python bindings. When in the latter "mode" it needs to be passed differently constructed objects through RMMNumbaManager.memalloc. Numba expects to deal entirely with one or the other situation at once, so if RMM doesn't "follow", there will be various internal breakages when trying to launch numba kernels.

Steps/Code to reproduce bug

The following should be run with NUMBA_CUDA_USE_NVIDIA_BINDING=1:

import rmm
from numba import cuda

cuda.set_memory_manager(rmm.RMMNumbaManager)
devary = cuda.device_array(shape=(3,), dtype="int32")

@cuda.jit
def f(x):
    pass

f.forall(3)(devary)

Should error with

ValueError: invalid literal for int() with base 10: b'\x00\x00\x00\xefE\x7f\x00\x00' 

Expected behavior
This kernel should launch and execute successfully, and do nothing.

Environment details (please complete the following information):

• Environment location: [Bare-meta]
• Method of RMM install: [source]

Additional context
relevant numba issue:
numba/numba#7929