Reconsider the design of cluster

[cupen]

pros:

1. It's really nice to find the address of an actor ( PID ) by Kind and ID, no matter how many cluster members.
   More specifically,, no matter how many processes running on how many machines.
2. grain codegen is good idea. It made a well-understand request-reply pattern which is enough to cover 95% of my need.

cons:

1. Too many IdentityHandoverRequests when cluster members scaling up or down, its time complexity is O(n^2 * (m/n)).
   n = How many cluster members
   m = How many actors totally in all of cluster members
   m/n = Average of how many actors in a cluster member
2. The ActivationRequest will spend 1 or more RTT, beacuse PartitionIdentiyActor will forward it serveral times in some worse cases. And also there is a worsest case (a->b->a->b) may causing high cpu usage.
3. The ActivationTerminated, its problem as above.