Encode JSON encode on write rather than read

[robacourt]

JSON encoding on write rather than read to reduce memory footprint. For example the memory use of an initial sync of a 35MB/200k row table can be reduced from 50MB to 25MB on the first initial sync (which includes the encoding and the writing to storage) and to 6MB on subsequent initial syncs (where we just read from storage and there's no encoding).

This change also allows some simplifications, so I have included them as refactoring in this PR:

• Logic to do with the structure and how to create Log Items has been consolidated to a new module LogItems
• The prepared_change intermediate state has been removed. The transformation Changes > list(prepared_change) > LogItems now is simply Changes > LogItems which makes the code more readable ( prepared_change was a 5 tuple rather than a nicely labeled map) and easier to reason about (since there's one less data structure to worry about)
• The snapshot row intermediate state has a few less references, ideally I'd like to get encapsulate it into a single module, but for now it's mainly Shapes.Querying that creates it and LogItems.from_snapshot_row/4 that reads it.
• Some logic duplicated in InMemoryStorage and CubDbStorage has been consolidated.

I've kept the refactoring in a separate commit to the functional change to aid with reviewing this PR.

[KyleAMathews]

Reminder to move this PR to electric — I'm going to close all the PRs tomorrow 🙏