general overhaul? no LXPLUS7, use CVMFS instead of AFS etc?

[jmuf]

These docs feel a bit dated - if they are still relevant, perhaps they should be updated:

• LXPLUS has (only) python3 nowadays (docs seem from LXPLUS7 times?)
• installing multiGB frameworks via virtualenv/pip/conda into AFS is a good recipe for atrocious performance on a single box, and an overloaded AFS if the trained model is used later e.g on LXBATCH. EOS is slightly better, but ideally would refer to some CUDA-compatible LCG release(s) from CVMFS that contain common libraries, and/or CMS-approved container images from some (scalable) registry.
• perhaps also refer to https://ml.docs.cern.ch/ ?