RlpMemo should work on compressed data

[Scooletz]

Currently, RlpMemo needs to Decompress data first to work on it. Then, it needs to Compress it before applying on the database. It could be cheaper and efficient if it could work on compressed data directly, and copy them over when one of the children is deleted/added and the underlying RlpMemo structure requires amendments (adding or removing 32 bytes when nodes are deleted added).

The cost of Compress and Decompress is paid by various components.

Decompress is used by ComputeMerkleBehavior when it reads a Merkle Branch that was not previously modified during a given block and needs to be amended. This is now partially amortized by the Prefetcher component that copies the data locally calling the decompression when it's needed. This helps a lot, but if we could ditch the decompression altogether we could either speed up the Merkle Prefetcher or abandon it altogether (smaller cost to be paid by merkleization that for storage is highly parallel).

Compress is called whenever the data are applied on the database. This is done by InspectBeforeApply and right now is not in the hot path of the execution. It's called by .FlusherTask so does not introduce an immediate penalty. There's a pending PR though #417 that moves it to the hot path, by calling .Apply method in the main thread. Whether its whole execution or at least a part of it can be offloaded is a subject to discuss, having no Compress call at all, would not make it slower.