Write-path scalability

[balegas]

Once we reduce the compute overhead (#1746), the bottleneck should move to disk IO. We're quite unoptimized right now, as we're fsyncing to disk quite frequently and data isn't downstreamed to clients while transactions are not persisted.

This task tracks the work necessary to be able to improve IO performance to saturate disk capacity without compromising the ability to downstream changes to clients.

Related issues:

• Send data out immediately on the write-path #1760
• Add a SQLite backed storage implementation #1742