[FEA]: Test nvJitLink and nvFatbin against sm_90a, sm_103 and sm_103a #3217

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

cuda.bindings

Is your feature request related to a problem? Please describe.

The architecture lists in the nvJitLink and nvFatbin tests stop at sm_100 and contain no arch-specific (a) target:

• cuda_bindings_13/tests/test_nvjitlink.py
• cuda_bindings_13/tests/test_nvfatbin.py
• cuda_bindings_13/tests/legacy_api/test_legacy_nvjitlink.py
• cuda_bindings_13/tests/legacy_api/test_legacy_nvfatbin.py

So nothing exercises the a targets, or sm_103 (B300 / GB300). cc 10.3 is where nvJitLink stopped serving sm_103 from sm_100 code, which is why #2296 moved the RDC fixture to -arch=all.

Describe the solution you'd like

Add sm_90a, sm_103 and sm_103a to the four lists. In the nvJitLink tests the PTX versions are paired with the architectures, so the new entries take each target's minimum PTX ISA version (8.0 for sm_90a, 8.8 for sm_103 / sm_103a). The nvFatbin tests cross an independent PTX-version parameter with every architecture, so their version list stays as is.

Describe alternatives you've considered

None; this only extends existing parametrize lists.

Additional context

I have the change ready and have run it on a B300 (cc 10.3) and a GH200 (cc 9.0): 186 → 282 passed on both, 0 failed, and sm_103a / sm_90a images produced through nvJitLink, NVRTC and nvFatbin load and run. I'll open the PR referencing this issue.