SciKitLearn python backend update breaks GH testing

[odunbar]

Old working version saw conda install
scikit-learn 1.1.1 and scipy 1.8.1
Now conda installs
scikit-learn 1.1.2 and scipy 1.9.0
I suspect the syntax for pyimport has changed

[tsj5]

This is puzzling -- the reason for the broken test appears to be that julia gets linked against one version of libstdc++, while the scikit-learn stuff wants a different version, as described here. The workaround (here) is to modify LD_PRELOAD; apparently there's no less brittle fix, since the issue is with the third-party scikit-learn dependency.

Aspects I'm unclear on:

• Is it necessary to use the LD_PRELOAD hack for any user of the scikit-learn backend on linux (seems likely)? It would be simple enough to modify LD_PRELOAD as part of the setup in Tests.yml, but this wouldn't be feasible to ask an end user to do every time they ran CES.
• What specific change to Manifest.toml brought in the incompatible scikit-learn? Conda.jl appears to pass through version specifications, so we could say julia --project -e 'using Conda; Conda.add("scikit-learn=1.1.1")', but this means we'd be stuck maintaining a compatibility matrix between julia versions and scikit-learn versions.

[tsj5]

Pinning the versions of scikit-learn and scipy appears to get the linux test to pass (PR #177). Should we proceed along these lines?