Strategy for reducing write amplification when handling random updates to multiple Segments

[simerplaha]

Currently we do not have a strategy to reduce write amplification when handling random updates that spans over multiple Segments.

Need to test the following approaches

1. Creating temporary levels where large spanned updates can occur gradually. Cost = space amplification and read overhead.
2. Allow level0's size to grow further. Costs = memory overhead. This is already configurable using Accelerator.
3. Allow un-merged segments to act a single unit. Similar to approach 1 but might have slightly lesser read overhead.