[FEA] Serialization of libcudf classes and exposing implementation details

[pentschev]

Is your feature request related to a problem? Please describe.
For multi-gpu polars, we will require serializing certain data in Python to be passed between Dask workers, for example aggregations. In #17469 I've proposed a way to do that, however, that proposal requires certain implementation details from aggreation.hpp,
more specifically classes derived from aggregation, such as std_var_aggregation. @vyasr has pointed out to the fact that those details are not exposed to pylibcudf and would be best if it continues like that.

Describe the solution you'd like
The solution proposed in #17469 seems to be the lowest hanging fruit, but as described above may not be considered optimal for several reasons.

Describe alternatives you've considered
Exposing attributes of the classes publicly may be an alternative, but that would incur in a different set of potential issues.

I'm not familiar with most of the design and options available in libcudf, so it's likely core developers will see other potentially better alternatives.