Idea board for how model uploads should run here

[martinkim0]

Necessary:

• Model upload workflows should run for every release of scvi-tools with appropriate revision tags

Nice to have:

• "Dry runs" of upload workflows with nightly or weekly scvi-tools to make sure newest changes are compatible
• "Mirrors" of tutorial models so that users have the option to set a (non-default) flag on the top of their notebooks to just use the pre-trained model instead of training themselves
• Tensorboard logs instead of storing losses in the model itself for easier viewing
• PPC metrics and/or integration metrics pre-computed

Not sure about:

• Whether we should have backport branches

Implementation:

• Workflows most likely have to run on self-hosted runners due to RAM limits and GPU acceleration for the subset of models that are trained on our end
• Need a new Docker image tag for these workflows
• We should be able to pass in revision tags as an argument to the workflow