BUG, ENH: np._from_dlpack: export correct device information

[charris]

Backport of #21119.

fixes gh-20340

It is possible to export incorrect device information if a new view has been created from an imported array:

import numpy as np
from torch

x = torch.arange(10).pin_memory()
y = np._from_dlpack(x)
y.base  # PyCapsule named 'numpy_dltensor' (contains device information)
z = y[::2]
z.base  # y (doesn't contain original device information)
z.__dlpack_device__()  # (1, 0) --> wrong (should be (3, 0) for pinned memory)

This PR fixes this behavior by walking the bases (if they are ndarray) as suggested in #20340 (comment). The test just coves the new changes but doesn't really test against an array from a different device.

cc @seberg @mattip