Use Divide-and-Conquer Algorithm in Scatter

## Enhancement Task
After split and before scatter, the split Regions are on three stores and the leaders are on one store. When scatter is executing, the TIKV with leaders will send a large number of snapshots, and this TIKV will become the bottleneck.

We can impl an optimization by first executing a coarse-grained scattering to disperse the leader. This step is to breaks down the scatter process into more sub-processes. And then executing a fine-grained scattering.
For example, if we need to split and scatter 10K regions in a cluster with 10 TiKVs, we can split into 10 regions first and scatter them. And then split and scatter the 10 regions.