AutoAnnotation run model script requires 'secret_keys'

[benhoff]

Due to some of the logic in the __init__.py module in the AutoAnnotation app, it currently requires that Django be fully installed, specifically for the secret keys.

See #516 for more details.

[benhoff]

Just thinking out loud, could look at moving the code into the engine application. Could put try catches around the relevant logic in the init file. Could also combine the logic from inference.py into model_loader.py and then do some Python path manipulations directly in the run model script so it doesn't touch any cvat logic.