fix(cuda.core): preserve DLPack C exchange errors

[rluo8]

Description

The DLPack C exchange API requires export/import callbacks to return -1
with a Python exception set
(dlpack.h;BufferError when the data cannot be described).
StridedMemoryView returned -1 without restoring the exception.

This PR restores the exception with PyErr_SetObject before returning -1,
adds regression tests for both export entry points and unsupported-
device import, and add a release note.

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-2939/
https://nvidia.github.io/cuda-python/pr-preview/pr-2939/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-2939/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-2939/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[Andy-Jost]

Thanks for the fix, and I confirmed it works. I think we can avoid adjusting the exception state by hand through CPython calls if we declare the Cython entry points except -1. The typedefs in _dlpack.pxd need the same annotation, or the assignment into the API table does not compile. That also covers KeyboardInterrupt, which except Exception lets through. I worked up an example in Andy-Jost@dbc22cf, and your tests pass unchanged on it. Take it or leave it, and I will approve either way.