coverage: add cuda.core tests for host, launcher cluster, strided layout, program-cache, and DLPack utils

[rluo8]

Description

Tests covered:

File	New tests	Target source
test_host.py	repr for all three Host forms; pickle/copy round-trip preserves the singleton via __reduce__	_host.py
test_launcher.py	thread-block-cluster cc gate (accept ≥9.0 / reject <9.0, Device mocked); cluster branch of _to_native_launch_config	_launch_config.pyx
test_program_cache.py	FileStream size-cap __delitem__/clear, entry-path iteration with missing/stray dirs, tmp-size accounting, touch_atime fallback + swallowed open/fstat failures	utils/_program_cache/_file_stream.py
test_strided_layout.py	validation/error paths: dense_like, repr-with-offset, min/max offset, broadcast_to, unsqueezed, repacked pack/unpack, max_compatible_itemsize	_layout.pyx
test_utils.py	as_tensor_map kwargs assembly; DLPack export round-trip over all NumPy-native dtypes + scalar/zero-volume shapes; versioned/unversioned capsule deleters; from_dlpack TypeError→unversioned fallback; CAI device-pointer/stream-ordering; DLPack C-exchange-API helpers	_memoryview.pyx

Coverage impact

Measured with paired full-coverage runs (no-patch baseline vs this patch):

Scope	Before	After	Δ
cuda/core/ overall	79.81%	81.84%	+2.03% (−213 missed)
_memoryview.pyx	65.82%	85.73%	+19.92%
_host.py	83.64%	100.00%	+16.36%
_launch_config.pyx	77.46%	91.55%	+14.08%
_layout.pyx	89.59%	96.23%	+6.64%
_file_stream.py	84.51%	90.49%	+5.99%

Testing

• Full local run on py314: all added tests pass.
• Cross-platform on local GPU runners: Linux x64 + Windows x64, green.
• Additionally verified on a B100 windows + Tesla T4 host.

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-2168/
https://nvidia.github.io/cuda-python/pr-preview/pr-2168/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-2168/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-2168/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[rwgk]

I glanced through manually and asked GPT-5.5 to do a full review. This looks solid to me. Great to see that our coverage numbers are getting significantly better.

[mdboom]

@rluo8: Thanks for this. It's really nice to see coverage improve.

These all look fine to me except the ones in test_utils.py. Rather than holding the rest of this up, could you remove those changes, and the submit test_utils.py alone as it's on PR? I'll comment on what I think could be improved there.

[leofang]

FYI, I mentioned in our team sync today. _StridedLayout is currently not exposed to the public, so all of these complexity is not touching the public API surface. In the near future we'll need to discuss if we still want to keep this object in cuda.core since in addition to maintenance burden it also adds to performance overhead (of creating StridedMemoryView), and if we decide to remove it all of these tests will be deleted right away.

As a result, I would recommend to not expand the tests here, although I don't feel strongly (since this file is already awfully long, before this PR).