support CUDA 12 sm_50 PTX

[harishappana-git]

CUDA 12.8 generates a few PTX forms that GPUOcelot did not understand. This updates the parser to accept the newer .loc and .debug_str syntax and builds CUDA tests for compute_50.

It also fixes the immediate operand types used by bfi, so valid CUDA-generated instructions pass validation and execute correctly. Regression tests cover the new PTX syntax, valid bfi execution, and invalid operand widths.

Tested with tinygrad using CUDA 11.5 / sm_35 and CUDA 12.8 / sm_50, on both PTX and NV mock backends: https://github.com/harishappana-git/gpuocelot/actions/runs/33059853145

[harishappana-git]

Hi Maintainer,

I’ve verified the current PR head with tinygrad using CUDA 11.5 / sm_35 and CUDA 12.8 / sm_50, on both the PTX and NV mock backends. The full validation run passes: https://github.com/harishappana-git/gpuocelot/actions/runs/33059853145

Please let me know if you’d like any changes here. Once this is merged and included in a new GPUOcelot release, I’ll update tinygrad to use that release and verify the corresponding CI changes there. Thanks