[CI] Bump CUDA requirement to 12.9

[trivialfis]

For a newer version of CCCL.

Also, the current Windows CI is still using 12.5, which is lagging behind the Linux CI.

[trivialfis]

cc @hcho3 is there any documentation about how to pack and update the Windows VM? I have access to a windows machine and can take the maintenance if there are some instructions/guidance.

[RAMitchell]

This is a moderately aggressive verion bump. 12.9 is from May 2025 I think? Just wondering is it practical to use CCCL independent of toolkit version?

This would only affect users building from source, but I could see it affecting some companies with restrictions on updating.

[trivialfis]

We can require an independent CCCL in cmake if that's needed.

Please note that, the requirement here is a newer user space program (nvcc, cccl, etc), not the GPU driver. You can continue to use an old driver (say , 570) with a new CUDA toolkit as long as it doesn't cross major version boundary (12.x). You can use the NVCC from conda NVIDIA channel.

I'm open to suggestions.

[RAMitchell]

An independent cccl version might give you more freedom to adopt new features with less disruptive build changes going forward. It might come with its own problems, I am not sure :)