Fairness meets Cross-Domain Learning: a new perspective on Models and Metrics