Support directly deploying Chronos endpoints without running a training job

[shchur]

Currently, to deploy a Chronos endpoint, the user needs to create a training job (essentially a no-op) before being able to deploy the endpoint. It would be much more convenient if the users could directly create endpoints or submit batch prediction jobs without requiring a separate training job.

This functionality could also be extended to other pretrained models like TabPFNMix used in the TabularPredictor.