split learning and inference methods and introduces new traits

[drahnr]

This would potentially allow gains when learning since dual numers/autodiff could be employed more effectively.