[Task] Create a protocol for common methods across model frameworks

[oliverholworthy]

Part of #669
Relates to NVIDIA-Merlin/Merlin#545

Background

Based on the discussions around having a common API across the backend framework implementations. We would like to make this concrete with a protocol that each model can be checked against to help make it clear which methods are part of a common API and which are specific to the various model implementations.

Description

Write down the common API methods supported by each model in a protocol that can be runtime checked. This can be extended in future. For the purposes of this task. The following may be enough to consider this task complete:

• training methods:
  • fit Fit (train) the model on a dataset provided
  • evaluate Evaluate the model on a dataset provided (returns metrics)
  • predict Run inference on the model returning the output predictions of the model
• io methods:
  • save Write out artifact(s) that can be reloaded by the load method
  • load classmethod to re-load a model from a previously saved model instance.