Core refactors

[mikeknep]

No user-facing changes here, strictly an internal refactor.

Previously, tables with JSON columns got "extra special treatment": after creating invented tables from them, we added the invented tables to the RelationalData graph as nodes, but not the source table itself; instead, metadata about that original table was stored in a RelationalJson object responsible for knowing everything necessary to transform invented tables back to the original shape. Then the RelationalData stored a dictionary of those objects. This was always a bit clunky and added a lot of mental overhead when trying to reason about the internal state.

In the spirit of "cattle, not pets", this PR makes what we now call "producer" tables more similar to normal and invented tables. Probably the key takeaway is that every table is a node in the graph, whether it is a "normal" table, an invented table, or a producer of invented tables. All state for a given table is stored on the graph node. This brings several benefits:

• the JSON module, with all its (de)normalizing transformation functions, is now stateless—we store all the table-mapping information on the producer table graph node
• backups no longer need to care about this extra collection of objects (RelationalJsons)—they just write down everything known about all tables, and (by extension) the restore process is simplified. (RelationalJson backup was also a bit "lossy"; no more!)
• elevating the concept of "producer tables" helps establish a more sustainable pattern for if we ever invent tables for other reasons in the future

There are also a few internal helper methods added to generally make things more readable.