cuda.core: emit -numba-debug to libNVVM, not --numba-debug

[fedonman]

Summary

Closes #2570.

libNVVM accepts only single-dashed options, but _prepare_nvvm_options_impl emitted --numba-debug, so every ProgramOptions(numba_debug=True) compile on the NVVM backend failed with NVVM_ERROR_INVALID_OPTION. numba-cuda emits -numba-debug for the same path. The NVRTC backend tolerates both spellings and is left alone.

The guarding test never caught this because it probed libNVVM with the same wrong spelling, so it skipped on every configuration, including CUDA 13.3. It is one of the entries in #2077.

Changes

• _program.pyx: emit -numba-debug in _prepare_nvvm_options_impl. The NVRTC call site is unchanged.
• tests/test_program.py: probe and assert the emitted spelling, add an assertion that the double-dashed form is absent, and drop the skip reason's claim that the option needs CTK 13.2 (libNVVM rejects the double-dashed form at any version).
• docs/source/release/1.2.0-notes.rst: release note. The wrong spelling is in released 1.1.0 and 1.1.1.

Verification

test_nvvm_program_numba_debug previously skipped everywhere. It now runs and passes, and it fails without the production change:

$ git stash push cuda_core/cuda/core/_program.pyx   # production only
$ python -m pytest tests/test_program.py -k numba_debug -v
tests/test_program.py::test_nvvm_program_options_as_bytes_numba_debug FAILED [ 50%]
tests/test_program.py::test_nvvm_program_numba_debug FAILED              [100%]
cuda.bindings.nvvm.nvvmError: ERROR_INVALID_OPTION (7)
NVVM program log: libnvvm : error: --numba-debug is an unsupported option

With the change:

$ python -m pytest tests/test_program.py -k numba_debug -v
tests/test_program.py::test_nvvm_program_options_as_bytes_numba_debug PASSED [ 50%]
tests/test_program.py::test_nvvm_program_numba_debug PASSED              [100%]

Whole module, before and after:

$ python -m pytest tests/test_program.py -q   # upstream/main
110 passed, 2 skipped in 2.01s
$ python -m pytest tests/test_program.py -q   # this branch
111 passed, 1 skipped in 2.61s

[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.