fbgemm_gpu not building correctly on CentOS 7 pytorch/manylinux-builder:cpu containers

[atalman]

We are seeing this failure at the end of the nightly Linux build smoke test:
https://github.com/pytorch/FBGEMM/actions/runs/12049477907/job/33621272179

WARNING:root:Could not the library 'fbgemm_gpu_py.so': /lib64/libstdc++.so.6: version `GLIBCXX_3.4.29' not found (required by /__w/_temp/conda_environment_12049477907/lib/python3.9/site-packages/fbgemm_gpu/fbgemm_gpu_py.so).  This may be expected depending on the FBGEMM_GPU variant.

Smoke test is green but this is actually hard error. Manifest in torchrec here:
https://github.com/pytorch/torchrec/actions/runs/12049473108

Looks like after switching to pytorch/manylinux-builder:cpu containers this does not work correctlty. Previous nightly run was not showing this error: https://github.com/pytorch/FBGEMM/actions/runs/12029739812/job/33535665003

Reference PR switching to these containers: pytorch/test-infra#5986

[q10]

Hi @atalman, due to recent updates in PyTorch nightly, we had to upgrade the version of gcc we used for building FBGEMM, and by extension, the minimum version of glibcxx had been updated - see the comments added to the code in #3398 for details.

The issue can be worked around by building and running FBGEMM inside a Conda environment, where one can control the glibcxx version that is being used - see https://pytorch.org/FBGEMM/fbgemm_gpu-development/BuildInstructions.html#fbgemm-gpu-build-setup-env for details.