Problem: memiavl persisted node queries is not optimal

[yihuang]

Leaf Separation

The idea is store leaf nodes separately, so we can directly do binary search in it, using the index.

Traversing the iavl tree is similar to binary search, but with directly binary search on leaf nodes, we have less indirections, based on local benchmarks, this optimization could reduce the time of Get from 200+ns(little endian byte order) or 170ns(native byte order) to 100ns, faster than the performance of MemNode.