[Take 2] Setup TorchBench in Docker

[huydhn]

Fix and reland #158613, I keep checkout_install_torchbench in .ci/pytorch/macos-test.sh script because it's still used there, and there is no Docker.

Testing

MacOS perf nightly run https://github.com/pytorch/pytorch/actions/runs/16580798470

cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @jataylo @hongxiayang @naromero77amd

[ZainRizvi]

Diff since take 1 looks reasonable. I trust you to make sure whether the docker failures are relevant or not

[huydhn]

cc @atalman I'm pretty sure the Docker image validation failures is not related, but not sure why it shows up in PR. Maybe that workflow is not meant to be run on PR (@XuehaiPan added the ciflow/nightly label to the PR)

[huydhn]

@pytorchbot merge -i

[pytorchmergebot]

Merge started

Your change will be merged while ignoring the following 7 checks: Build Official Docker Images / validate / cuda12.6-cudnn9-runtime / cuda12.6-cudnn9-runtime, Build Official Docker Images / validate / cuda12.6-cudnn9-devel / cuda12.6-cudnn9-devel, Build Official Docker Images / validate / cuda12.8-cudnn9-devel / cuda12.8-cudnn9-devel, Build Official Docker Images / validate / cuda12.8-cudnn9-runtime / cuda12.8-cudnn9-runtime, Build Official Docker Images / validate / cuda12.9-cudnn9-devel / cuda12.9-cudnn9-devel, Build Official Docker Images / validate / cuda12.9-cudnn9-runtime / cuda12.9-cudnn9-runtime, Build Official Docker Images / validate / cudacpu-cudnn-runtime / cudacpu-cudnn-runtime

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging

Check the merge workflow status
here

[huydhn]

cc @anijain2305 I'm re-landing this PR today, so I suspect that you might want to rebase #159291 before landing it. There shouldn't be any major change, this PR just move TorchBench installation during test time into the Docker build time.

[jithunnair-amd]

@huydhn This PR seems to have had a collateral impact on the only numpy2-based workflow that runs GPU tests: https://github.com/pytorch/pytorch/actions/runs/16763863956/job/47469669670

[huydhn]

@huydhn This PR seems to have had a collateral impact on the only numpy2-based workflow that runs GPU tests: https://github.com/pytorch/pytorch/actions/runs/16763863956/job/47469669670

Yes, you're right. @izaitsevfb has notified me earlier about the same issue. As this PR has lots of touch points across multiple workflows. I propose that we forward fix it in #160047. Basically, the regular ROCm CI job doesn't need to include the baggage from TorchBench.