Implement asynchronous snapshot by copying dataTree

[lzydmxy]

Which issues of this PR fixes:

This PR is part of #144

Change log:

• Creating snapshot asynchronously.

Motivation:

Right now keeper creates snapshot synchronously, during which period, it is unable to handle to user requests for creating snapshot is in the request handling procress.

If a cluser has large amount of data, creating snapshot will lead large blocking time. And because keeper can not handle heartbeat, leader election may happen. Both of them will make keeper unavailable.

Solution:

The main idea is a little like COW.

1. Copy all nodes of the data tree (copying is parallel) and save them in memory temporarily. This step will block user requests and consume some memory.
2. Serialize and persist the data, this step is in background.

Optimization result:

We have a keeper cluster which holds metadata of Clickhouse and it has 58659723 nodes.

1. The blocking (user requests) time reduce from 180 seconds to 4.5 seconds, it is about 2.5% of the orignial.
2. But there is also a side effect. When creating snapshot memory usage increases from 40GB to 60GB (approximately 1.5 times).

[JackyWoo]

@lzydmxy nice work, please show us the memory usage when creating snapshot.

[JackyWoo]

I will carefully review this PR tomorrow.

[JackyWoo]

@lzydmxy LGTM except some code sytle issues.