Lint for variable scoping serialization issue? (Or just fix it?)

[cisaacstern]

By far the most common Dataflow bug we are seeing on staged-recipes is the serialization issue where module-level names are not available within function scopes. A few examples from the last 24 hrs:

• NWM 1km LDAS rechunked dataset staged-recipes#233 (comment)
• C-iTRACE tracer pointing to NetCDF4 files staged-recipes#230 (comment)
• AWS NOAA Optimum Interpolation SST staged-recipes#232 (comment)

Maybe we need a custom linter to flag this before we /run the tests?

Even better would be to fix it, if possible. I believe the Beam option save_main_session (which we use) is supposed to address this, but it doesn't appear to work, and some knowledgable folks at the Beam Summit I attended in Austin last July told me this was always a "hacky workaround". If I remember correctly, I was told using a setup.py as described in the Beam docs here might possibly resolve this. I'll do an experiment to see if that's the case.