CUDA: fix PDL CC check for JIT compilation

[JohannesGaessler]

See discussion starting at #22522 (comment) .

On master we are currently checking the raw CC value vs. Hopper for PDL. However, what we should be doing is use ggml_cuda_highest_compiled_arch(cc) for the check instead to avoid a mismatch between host and device code for JIT compilation.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: No

clicked a wrong button

[ORippler]

Had the time to read up on ggml_cuda_highest_compiled_arch now, thanks for fixing this

[ORippler]

@JohannesGaessler Just a heads-up, but I think we will be generally facing the same issue for family/architecture specific features that we have for mxfp4/nvfp4 acceleration on BW currently (i.e. compilation failing for sm_120 as we have no way to distinguish between family/arch specific compilation on host code and us having to monkey-patch it to sm_120a in cmake). On this virtue, it actually makes sense to compile for sm_90-virtual, as this will guarantee correct behavior for PDL.

To explain:

should a user compile only for sm_90a, __CUDA_ARCH_LIST__ will include sm_90, yet the sm_90a ptx will not be forward-jittable. Our host code will think we can dispatch to PDL on say sm_120, while the device code from sm_89 does not contain the required barriers.

I'll try to push again for including arch/family specific compilation into __CUDA_ARCH_LIST__/onto host-side (it currently is available only in __CUDA_ARCH__), or at least ask for a robust work-around.