Support for __partitioned__ as data source

[fschlimb]

An API/protocol for exchanging distributed, partitioned data is being defined. The goal is to establish a uniform way of describing such data while making it possible to avoid unnecessary data copies and even zero copy and so avoid specialized implementations for different data containers (like RayDataSet, MLDataSet, modin etc). The protocol basically defines the "meta" data about the distribution and partitioning (tiling) of the data. The current proposal is here: https://github.com/IntelPython/DPPY-Spec/blob/draft/partitioned/Partitioned.md, an issue/discussion is here: IntelPython/DPPY-Spec#3

Any feedback/input to the discussion/spec is welcome.

There are several implementations/PRs for several projects, like HeAT, modin, MLDataSet and DAL.

It would be nice to have xgboost_ray support this, too. I will also open a PR.