Batch Inference off of previously trained (and saved) model

[johnml1135]

We now have the ability to save models from #268. There are users that may want to do a quick inference off of a previously saved model, (rather than retraining). This would reduce the time to get new drafts from 8 hours to under 15 minutes (5 minutes?). This would require:

• API design in Serval to indicate "just inference, don't retrain"
• Scripture Forge work to choose "just inference, don't retrain"
• Machine.py updates to enable referencing the previously saved model

API design

• The Nmt engine is marked for persistence.
• For build options:
  • overloard parent_model_name so that if it is set to saved
  • set train_params.do_train to false.
• Choose new queue that includes the 3090's for fast inference - needs: New NMT option - choose ClearML queue machine#151

SF design:

• Some form of toggle to choose previous build or do training afresh (and communicate 8 hours vs. 15 minutes)

Should it be auto-magical?

• What about a pop-up to the user in SF that no (or minimal) training text has changed - do they want to just inference?