🐛 Bug

There's regularly issues building PyTorch with either conda compilers or with a very new (system installed) GCC + conda Python+dependencies. Using a development setup based on conda is recommended (https://github.com/pytorch/pytorch#from-source) so this is not a great situation. Example issues:

• error: 'SO_REUSEPORT' was not declared in this scope #29093 (has some relevant discussion)
• pytorch build fails: __NR_getcpu undeclared, a possible fix included cpuinfo#36
• The patches we're carrying around for a build setup used by ~8 people, gets quite inefficient: https://github.com/Quansight/pearu-sandbox/blob/master/working-envs/activate-pytorch-dev.sh

Furthermore, making the build config with conda compilers robust is required for making the conda binaries conda-forge compatible so they can be synced to conda-forge/pytorch (there's more to do there that I need to follow up on, this is step 1 - xref conda-forge/pytorch-cpu-feedstock#7).

Expected behavior

Using conda compilers on Linux should work out of the box. To make it more reliable, we should add a CI job using it. The question I have is if it should be an extra CI job entry, or if we should modify an existing one (and if so, which one)?

Cc @ezyang

cc @malfet @ezyang @seemethere