Default `stage` for transforms (especially `Standardize`)

[vpratz]

While using BayesFlow, I have managed multiple times to ruin my trained adapter when trying to use it directly, as the naive adapter(data) call defaults to stage="training" for the Standardize transform. I would propose changing this, as I think it is a footgun that can mess up the pipeline/results after training without the user noticing necessarily. It's also not obvious that one has to be on the lookout for this, as the momentum is enabled by default.

@LarsKue @stefanradev93 What do you think?

[stefanradev93]

I agree. Let's align this with all keras default behavior, which is training=False by default. Could you make a quickfix PR?

[vpratz]

Closed by #418