[Pytorch Upstream][Regression] Pytorch CI build triton wheel failed because of out-of-memory.

[etaf]

Describe the bug

After upgrade triton commit to c4a79a1 , we find the pytorch CI build triton wheel failed because of out-of-memory.

https://github.com/pytorch/pytorch/actions/runs/13492607455/job/37697482583?pr=147727

I've tried to set the MAX_JOSB=8, and then the build job can passe: https://github.com/pytorch/pytorch/actions/runs/13515462982/job/37763241311?pr=147727

Environment details

Triton: c4a79a1
Pytorch: bea72180ed75f522ce4fe5e723bc2112e0874732
OS: Linux
GPU: PVC

[alexbaden]

As discussed offline let's use MAX_JOBS for the PyTorch release.
I think oneAPI 2025.1 should improve the situation somewhat. We can also look at LLVM linking order - it will be interesting to see if the NVIDIA/AMD backends start having issues when they update their pins.