This PR may be controversial, but I noticed our Docker container now uses Python 3.8, and hence all of our tests are now only running on Python 3.8 (from what I can see?).

• (Edit: there's also a workflow I missed that does a conda install to Python 3.7, but it doesn't run the test suite.)

My experience is that it's very easy to break 3.6 support by accident if not testing with it. And we're increasingly seeing dependencies drop support for 3.6 on other repositories in the ODC ecosystem.

Some options:

1. A recommendation of 3.8. The first commit updates all docs to recommend and default to Python 3.8, but setup.py still allows Python 3.6 installs.
2. A hard requirement on 3.8: the second commit requires Python 3.8+ in setup.py too, so users cannot accidentally pip install newer versions on 3.6. (edit: and to bump conda smoke test version)

I'm happy to remove the second commit if people think it's too harsh. But I think a minimum requirement is friendlier to users than exposing people to unexpected bugs.

Thoughts?

ps. The other option, of course, it to make 3.7 the minimum. But is that still common anywhere, among non-Docker users? And do we want to test with it?
pps. I also bumped the Postgres version recommendation, since none of us test on 9.5. But there's no hard requirement.