register_host_memory fails for numpy array created frombuffer

[amishatishpatel]

Bug description

Cannot pycuda.driver.register_host_memory using numpy array that was created frombuffer.

Error thrown

ValueError: Cannot set the NumPy array 'base' dependency more than once

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "/path/to/my_test_file/./test_pycuda.py", line 12, in <module>
    pycuda.driver.register_host_memory(array)
SystemError: <Boost.Python.function object at 0x1b80350> returned a result with an exception set

Browsing other issues raised, this might relate to issue #450 and PR #451.

• I see the corresponding unit test was marked xfail in 2a276c4

However, this is reproducible with numpy==1.26.2 (not numpy 2.0 or newer).

Steps to reproduce

Executing the following minimum working example with pycuda==2024.1.2 results in the above error.
Note that executing the same minimum working example with pycuda==2023.1 does not throw an error.

#!/usr/bin/env python3

import mmap

import numpy as np
import pycuda.autoinit
import pycuda.driver

size = 1024 * 2048
mapping = mmap.mmap(-1, size, flags=mmap.MAP_SHARED)
array = np.frombuffer(mapping, np.uint8)
pycuda.driver.register_host_memory(array)

Expected behaviour

I'd like this to be fixed, please. This is quite key in (our) benchmarking of GPU (htod, dtoh) performance,
if/when we do upgrade to the latest pycuda.

Environment

• OS: Ubuntu 22.04.4 LTS
• CUDA version: 12.5
• CUDA driver version: 555.42.06
• PyCUDA version: 2024.1.2
• Python version: 3.12.5

Additional context
Belated apologies in advance if I've misinterpreted the error/this is already on your radar. Figured I'd
flag it anyway, just to be safe.