Extensions can should not be allowed to impact global environment during builds

[Micket]

Example of this happening: The PyTorch bundle has torchvision and torchaudio.
https://github.com/easybuilders/easybuild-easyconfigs/blob/df6a3afb8ad03f0bc72c2652a3f3781c56af0b6b/easybuild/easyconfigs/p/PyTorch-bundle/PyTorch-bundle-2.1.2-foss-2023a-CUDA-12.1.1.eb#L104

torchvision sets the environment variable BUILD_VERSION globally, and torchaudio happens to have a setup.py that does

def _get_version(sha):
    with open(ROOT_DIR / "version.txt", "r") as f:
        version = f.read().strip()
    if os.getenv("BUILD_VERSION"):                                                                                       
        version = os.getenv("BUILD_VERSION")
     elif sha is not None:
        version += "+" + sha[:7]
    return version

which introduces the bug that the version is misreported for torchaudio as 0.16.2, despite it being 2.1.2

1. We should avoid setting environment variables globally, and going forward always stick to using run_cmd(..., env=custom_env).
2. Framework should record and restore the environment before and after each extension.

This goes to os.environ as well as the cwd.

I don't know if we have any situations where we rely on leaky behavior like this, though we really should avoid it.

[boegel]

Framework should record and restore the environment before and after each extension.

@Micket This is the correct fix, extensions shouldn't be influencing each others build environment, that's a bug...

[boegel]

@pavelToman Are you up for taking a look at this?

It probably boils down to recorded and restoring the environment before/after installing an extension, in install_extensions_sequential, you can use the restore_env function for that.

Not sure what needs to happen to also make that work correctly with install_extensions_parallel, since environment variables are global...