CFL: On the Use of Characteristic Function Loss for Domain Alignment in Machine Learning