[WIP] Partition-metadata tracking in Dask-DataFrame

[rjzamora]

Brief proposal for a performance-motivated metadata update in Dask-DataFrame.

NOTE: Although this proposal is distinct from a high-level graph or query-optimization system, it should make such a system much easier to implement! I say this, because we will still want to be tracking and managing the same kind of metadata in one place.

See also: dask/dask#9473

[rjzamora]

cc @jrbourbeau @mrocklin for viz

There is not much to review here yet, but I am starting to organize my thoughts on this, and I'm feeling somewhat confident that this would be a worthwhile effort. The rough POC wasn't very difficult to put together, and I think proper HLG/HLE-optimization would require us to do much of the same refactoring anyway (unless we choose to do high-level optimization in a completely new library, that is).