Simple prototype of deploying model pipelines from a config on Ray Serve.

- Requires that you're running on the Ray nightly wheels (for Serve CLI).
- "Models" are defined in serve_pipeline/__init__.py. They're currently just random number generators but could have any Python code filled in.
- Pipeline is defined in example_pipeline.json. The "class" field must be the import path to a class that's installed in the Python environment on the Ray cluster.

To run:
  pip install -e serve_pipeline
  ray start --head
  serve start
  python deploy.py example_pipeline.json 
  curl -X GET localhost:8000/api # Should return a random float in [0, 1).
 

Other notes:
- Deploying the pipeline (deploy.py) could be done from a remote machine via the Ray client.
- The model code definitions should be built into the Docker image that the Ray cluster is running.
- This can support all of the Ray Serve features - scaling each model, batching requests, using GPUs, etc.
- Error handling is really sloppy right now (e.g., it won't clean up resources if it fails halfway through), but we could make this declarative.