python version in prognostic_run image

[brianhenn]

The prognostic_run image builds with python 3.6.9, whereas the fv3net image uses 3.7. This leads to occasional failures, e.g., vcm.round_time doesn't work in the prognostic_run image because it depends on singledispatch which doesn't operate below 3.7. Update prognostic_run requirements to 3.7?

[nbren12]

This is on my radar. To upgrade from 3.6 we need to upgrade the images in fv3gfs-fortran to ubuntu 20.04. See ai2cm/fv3gfs-fortran#162.

[nbren12]

My nix PR #1078 would also fix this.

[nbren12]

However this does reveal a problem with our testing strategy. We run tests in python 3.7, and I noticed some other problems when trying to run the fv3fit tests in the prognostic run image.