Benchmarks as reference point for users

[610v4nn1]

Currently users have hard time to understand what can be achieved using continual learning methods instead of ad-hoc heuristics. It would be great to provide some benchmarks against some simple baselines tracking both accuracy and training time.

Suggested baselines: retrain from scratch, fine-tune on latch data chunk.