Implement something similar to sklearns request_metadata

[AKuederle]

Pipelines that call algorithms that require certain additional parameters passed to their action methods, at the moment just pass all available metadata. This is a little annoying...

We could build a way similar to the metadadta routing in sklearn (https://scikit-learn.org/stable/metadata_routing.html) to have algorithm request that additional information is passed to them.

The problem is, how do we standardize this. The most generic version would be that the algorithm has a method that can extract the requested metadata from the datapoint. However, this means algorithms suddenly need to know about datasets. The whole idea behind pipelines is to extract way the dataset on only pass the relevant information to the algorithms.

The other way around, we could just be more surgical and supplying algorithms only with the metadata the need, by having a way to understand what metadata an algorithm needs. This would not make the work for pipeline authors any easier, but it would allow us to raise better error messages, directly pointing to the lack of certain metadata available and each algorithm would only get information that they activivly request.

Not sure, if this provides a lot of value though....