Upgrade torch

[dxqb]

• Torch 2.11 is still affected by torch.compile VRAM usage regression between 2.9.1 and 2.10.0 pytorch/pytorch#175058
• Torch 2.11 requires torchvision 0.26. Torchvision video writing has been removed: https://docs.pytorch.org/vision/main/generated/torchvision.io.write_video.html - fixed here: Remove Depedency on TorchVision for video creation #1393
• is video reading in mgds also affected? - No

[dxqb]

Status:

• Torch is deprecating CUDA 12.8 which we are currently using
• they are keeping CUDA 12.6 as legacy for old devices
• so upgrading to CUDA 13.0 is necessary for the next torch upgrade
• CUDA 13.0 requires NVIDIA driver version >= 575
• with torch 2.13 torch.compile VRAM usage regression between 2.9.1 and 2.10.0 pytorch/pytorch#175058 will probably be fixed
• but with our current torch 2.9 there is another major issue [Bug]: Qwen Image sampling VRAM regression #1371

so we could wait for 2.13 or upgrade earlier and trade 1 bug for another one

[O-J1]

LGTM!

[dxqb]

as this PR also upgrades to CUDA 13:
barely any hosts on vast.ai are still on drivers that can't use CUDA 13. this has been a problem a while ago (see #1285) but not anymore it seems