Support CUDA 13

[LegNeato]

Ideally we detect and dynamically switch.

[LegNeato]

We'll need something like #227 as well as handling the API changes.

[PMQ9]

Hi @LegNeato, I kinda want to take this on, but I only have Turing, Ampere and Ada GPUs, I don't have any Blackwell. So if I take this on, I think I can try to get CUDA13 to work with all of the GPU architectures except Blackwell. I assume that should cover a lot of the GPUs people are using right now.

[LegNeato]

I'm currently working through it, there are some gnarly issues.

[LegNeato]

@nnethercote you got this working, correct?

[nnethercote]

I have it working locally with some hacks. I am working through turning the hacks into proper fixes. #311 is one example.

[nnethercote]

#312 also helps, by setting the default compilation target to compute_75, which is the minimum supported by CUDA 13.0.

[nnethercote]

#314 is another step, handling some of the API functions that changed their type signature.

[nnethercote]

#317 is another one.

[nnethercote]

All the above-mentioned PRs have been merged. Things are working on my ASUS Ascent GX10 box with CUDA 13.0 as well as they are on CI. I didn't hit any LLVM/NVVM issues. I think we can declare victory and close this issue.