Add Ray integration for Pipelines

[oryx1729]

Ray Integration

This PR is an initial prototype of how Haystack Pipelines can be scaled horizontally and distributed across a cluster of machines using the Ray framework.

Design

The design goal is to provide a standardized interface for Pipelines similar to the Haystack Components. A new BasePipeline class is added that is extended by Pipeline and RayPipeline. This allows users to switch between "execution engines" for the Pipelines while retaining the configuration for the Components.

The new RayPipeline has methods similar to the Pipeline class. With the current implementation, a RayPipeline can be instantiated locally or on a remote cluster using a Pipeline YAML configuration. Ray provides a handle, that allows easy integration of deployed Pipelines with Python code or REST APIs.

Ray allows independently scaling the components by adding more replicas. For instance, in a typical extractive QA Pipeline, multiple readers can be added in front of a single Elasticsearch instance. The number of replicas can be configured by the new replicas parameter for each component in the Pipeline.

Implementation

The nodes in a Pipeline are wrapped in individual Ray Deployments. The Pipeline.run() calls the "graph" of deployments to get the results for an input query(or indexes an uploaded file for indexing Pipelines).

To create a deployment, the component type is passed to Ray together with the YAML configuration. The deployment then creates an instance of the component using parameters from the pipeline config.

Breaking Changes

The current YAML configuration has a type attribute for pipelines that is being used to specify Query or Indexing pipelines. This conflicts with the type parameter for components, where it is used for component class.

With this PR, the type now determines the class to use – Pipeline or RayPipeline.

The pipeline_type parameter is now replaced by root_node.

[tholor]

Looks already pretty good to me! Great job!
Left a few smaller comments.
Also, I have one high-level question: Did you test how much overhead we create with the communication between ray nodes? So comparing the time of a run with the regular Pipeline vs. RayPipeline on a single, local machine.

[oryx1729]

Also, I have one high-level question: Did you test how much overhead we create with the communication between ray nodes? So comparing the time of a run with the regular Pipeline vs. RayPipeline on a single, local machine.

From Ray's documentation, the overhead should be in single-digit milliseconds. I haven't benchmarked with the Pipelines yet.

[tholor]

Looking good. Only more documentation is needed to ensure a lower entry barrier for users + collaborators.