Consider the NVRTC major version when deciding whether cubin is available

[LeSingh1]

        _, nvrtc_minor = nvrtc.version()
        use_cubin = nvrtc_minor >= 1

nvrtc.version() returns (major, minor) and the major is discarded, so the test is
really "is the minor at least 1". That is wrong for every x.0 release:

nvrtc.version()	nvrtc_minor >= 1	correct?
(11, 0)	False	yes
(11, 1)	True	yes
(12, 0)	False	no
(12, 9)	True	yes
(13, 0)	False	no

The intent is "NVRTC is new enough to emit cubin" — nvrtcGetCUBIN arrived in CUDA 11.1.
On a 12.0 or 13.0 toolkit the check silently selects
--gpu-architecture=compute_XX + nvrtc.get_ptx() instead of sm_XX +
nvrtc.get_cubin(), so the examples JIT-compile PTX at module-load time rather than using
the cubin they asked for. It is a silent fallback, not a crash — which is why it has gone
unnoticed. Every example here carries a cuda-bindings>13.2.1 PEP-723 header, so a 13.0
toolkit is a realistic configuration.

Both copies of the expression are affected:
cuda/bindings/_example_helpers/common.py:58-59 (used by KernelHelper, i.e. most
examples) and examples/extra/jit_program.py:71-72.

Fix

Compare (major, minor) as a tuple against (11, 1).

In common.py this becomes a named nvrtc_supports_cubin() helper, so the decision table
is unit-testable without an NVRTC install. jit_program.py does not import the example
helpers, so it gets the equivalent inline comparison plus a comment rather than a new
dependency.

I kept the 11.1 threshold exactly as-is — this PR is only about the major being ignored,
not about re-deciding where cubin support starts.

Test

New cuda_bindings/tests/test_kernel_helper.py covers the full table above, plus a case
that pins the exact disagreement between the old and new expressions for (12, 0) and
(13, 0). Pure Python — no NVRTC, no toolkit, no GPU.

ruff check and ruff format --check are clean on all three files.

[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.