Improve ML Container

[ax3l]

Follow-up to #195

• it seems that podman-hpc pull does not follow the logic we want: pull if a newer image was found, otherwise use last copy (it seems it always uses the local copy, even if a newer image was published) ML Batch Script: Pull Conditionally #303
• it turns out the PSCRATCH location is slow to start off, instead let us use again CFS and its DVS_RO location https://docs.nersc.gov/development/containers/podman-hpc/overview/#using-shared-images
  • PSCRATCH squash image has: 12-14sec import time
• we need to add a conda lockfile workflow as in the GUI, to make sure we only intentionally change conda-packages (for stability of builds)
• we need to add the ML image/container build to a script similar to publish_ml.py Add publish_container.py #309