fix(core): use the NVVM numba debug option spelling

[aryanputta]

Fixes #2570.

libNVVM accepts the single-dash -numba-debug option, while cuda-core emitted --numba-debug for the NVVM backend. The latter is rejected as NVVM_ERROR_INVALID_OPTION.

This change:

• emits -numba-debug for NVVM only;
• updates the capability probe so the regression test no longer skips because it probes the invalid spelling;
• updates the option serialization and compile-path tests;
• adds a release note.

Validation:

• Ruff check and format passed.
• Cython lint passed.
• Focused pytest collection was blocked on macOS because the CUDA test helper requires Linux libc.so.6.
• Lychee was blocked by sandbox network access to existing external links.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[lijinf2]

Thanks for the contribution!

This looks like a duplicate of #2571, which was opened by the issue author to address the same bug in #2570. Both PRs fix the NVVM backend to emit -numba-debug instead of --numba-debug.

[lijinf2]

@aryanputta Heads up for future contributions: please check the issue’s linked PRs / Development section first, in case a fix is already in progress.

Thanks again for taking the time to put this together — your help is appreciated!