Utilise stored posterior values during NF training

[thomasckng]

Currently, the NF training only cares about the sample density. We should explore directions to utilise the log_prob stored with each samples. Main categories of current idea include modifying the loss function, sample more on high log_prob region, reweight/resample from the sample pool for pick samples for training.