cold start performance

[jchris]

related to #125

this issue is to track the idea of segmenting the dataset so the reader can pull just the current read only set, the log, or even the hinted result of 1st paint, as a single GET.

From discord regarding v0.17 in March:

Performance seems suitable up to about 100k encrypted events in a database, at which point cold start time becomes the next optimization frontier.

If we output the Pail and the event log in different streams we can snapshot the Pail root for quick cold loading. We can also split the document bodies into their own stream. We'll want to do level-based compaction so we aren't compacting the whole thing as frequently, and we can add policy so that events outside of the finality window are left for archive but not synced.

With these improvements, we should expect the cold start performance of the database to be a function of the number of records, not the number of events.

The frontier will then become lazy loading of the CRDT root Pail, so big datasets can query before full load.