Implement bucket rebalancer

[Gerold103]

Rebalancer is started on one of cluster nodes, for example, on a master with minimal UUID, as it works in VShard. Rebalancer asks bucket count from different replicasets, calculates a perfect balance, and sends bucket routes to overflowed replicasets. Quite simple implementation.

Implement rebalancer routes applier. The most important and complex part of applier is dealing with bucket RW references. The problem is the same as in tarantool/vshard#173, but it is sharper, more rigorous since GC works on already dead buckets, not under heavy load, while rebalancing moves alive buckets, and it should labour as fast as possible. Possible solutions are use the same way as for RO refs, or use only in-memory locks, for example. A one could ask - how is it possible to have RW not on a master, applying routes? - the answer is: when you have multiple masters in a replicaset. For master-slave routes applier is trivial, but master-master aggravates all the things, tarantool/vshard#165.

Follows #3988.