DLPack-related tests & examples with host-accessible memory are sensitive to NumPy versions

Minimal reproducer I just built for nvbugs 5591656:

import numpy as np
from cuda.core.experimental import Device, LegacyPinnedMemoryResource

dev = Device()
dev.set_current()
mr = LegacyPinnedMemoryResource()
buf = mr.allocate(40)
arr = np.from_dlpack(buf).view(np.float32)
arr[:] = 1

This causes ValueError: assignment destination is read-only until NumPy v2.2.5 as Rui from the QA team correctly pointed out.

Previously, we've known NumPy's DLPack support is not as stable as we hoped. Some tests required v2.1.0+ to run. This seems to be a new situation that we did not capture so far, because we let most, if not all, of the test dependencies float.

We should re-evaluate how aggressively we'd like to teach users to use NumPy for host interoperability, and examine all the test-skip conditions.

Start with the minimal reproducer in the issue and locate the DLPack tests and examples covering host-accessible memory. Compare behavior across the NumPy versions mentioned, including the v2.2.5 boundary, then review every existing test-skip condition; done means the affected tests and examples have an explicit, reproducible version policy.