Automated normalization strategies for finite differences

[dionhaefner]

Summary

See #469 (comment)

Why is this needed?

see above

Usage example

see above

[andrinr]

Great call. Related, in the demo we also needed normalization to make gradient descent work over the pipeline. So possibly normalization strategies could be implemented in a way so they're usable not only for FD but also more broadly.