Run scatter region in parallel

[kennytm]

Feature Request

Describe your feature request related problem:

Currently "scatter" is run immediately after "split", all in a single thread. While "split" must be serial, "scatter" does not need to be.

Describe the feature you'd like:

Change (*RegionSplitter).Split() to:

1. Perform all split without any scatter
2. In parallel (with some concurrency limitation), perform scatter.

Describe alternatives you've considered:

Teachability, Documentation, Adoption, Migration Strategy:

[kennytm]

We already scatter in parallel to some degree after #167.

• The ScatterRegion only inserts a "scatter" operator, which is executed asynchronously. (This is also why we need to "wait" for the scatter operation to be completed in the end.)
• We perform the "wait" every 128 regions (the "128" is controlled by --concurrency, the count is before splitting).