Enable CUDA 12.9 binaries

[tinglvv]

🚀 The feature, motivation and pitch

Adding CUDA 12.9 to nightly binaries for new arch support

• Add CUDA 12.9.1 x86 nightly binaries #154980
• Add CUDA 12.9.1 sbsa nightly binaries #155819
• Add almalinux CUDA 12.9 docker build, required for magma build #155340
• Build magma-cuda 129 #155496
• [BE]: Sync cusparselt 12.9 with static build and other cuda 12 #155709
• Add CUDA 12.9 libtorch nightly #155895

domain builds (torchvision, torchaudio)

• add CUDA 12.9 domain builds for linux test-infra#6761
• Add windows 12.9 domain builds test-infra#6838

windows build

• Add CUDA 12.9 Windows AMI test-infra#6734
• Add Windows cuda 12.9.1 build #156630
• Add Windows CUDA 12.9.1 build #155748
• [CD] Windows Magma build 12.9 and cuda scripts #155799

Alternatives

No response

Additional context

No response

cc @seemethere @malfet @osalpekar @atalman @ptrblck @msaroufim @eqy @jerryzh168 @nWEIdia

[youkaichao]

hmmm how is it possible to build pytorch with cuda 12.9?

I think there's something wrong with cuda 12.9 , they removed targets/x86_64-linux/lib/libnvToolsExt.so, which torch.cuda.nvtx depends on.

there's a new file targets/x86_64-linux/lib/libnvtx3interop.so in cuda 12.9.0 and 12.9.1 , but i'm not sure if this file serves the same purpose.

I checked nvcr.io/nvidia/pytorch:25.05-py3 image too, it does not have targets/x86_64-linux/lib/libnvToolsExt.so either.