Use lighter runtimes when doing real-time inferences

[brnaguiar]

• https://www.advancinganalytics.co.uk/blog/2020/7/21/deploy-models-almost-anywhere-with-onnxai
• MLflow is deprecating MLeap runtime for ONNX.AI

[brnaguiar]

Won't do because this is a batch inference project, but we use Spark Pipelines with Custom Transformers, so Pipeline serialization to lighter runtime environments becomes trivial.