Implement adversarial training

[asogaard]

This can be used to mitigate parametrised systematic uncertainties or data/Monte Carlo differences. Having the ability to run this form of training seamlessly would be a big benefit.