Star Cubing alike partial cuboid flow down

Key idea

• Assume base cuboid fit in mem (for a mapper split of course)
• For child cuboid having common dim prefix, flow down records to child as soon as completed records can be fulfilled in child
• Keep completed records in mem as long as possible, but also save them to disk in the background
• When mem run short, pause flow down, remove completed records from mem, and resume flow from disk
• Multiple flow downs at the same time, configurable thread number