Kullback-Leibler divergence

[Atticus1806]

This is a draft for the Kullback-Leibler divergence (loss).

The idea is to use this with .mark_as_loss(). This is only a draft since I am not 100% sure about all the modalities, which might need to be taken care of (or are already dealt with with the marking)

In my mind this is:

• conversion to actual (batch-wise) loss
• potential inconsistencies / stability concerns with the input e.g. PyTorch (https://pytorch.org/docs/stable/generated/torch.nn.KLDivLoss.html) assumes estimated to be in the log-space already and makes it optional for target to be too. Do we want to enforce/handle this in the same way?

[albertz]

It should be consistent to cross_entropy regarding the logic of log-space etc.

Also see the discussion in #38.

I don't think it should also perform the reduction over time or other dims. It should just calculate the KL. The final reduction and accumulation over time and batch would be handled by RETURNN.

[Atticus1806]

Okay I added the checks for the type of estimated the same way it is done for ce. So I think this should be ready then.