feat: gossip protocol

[leiysky]

Part of #31

Summary

Introduced gossip protocol to manage cluster membership and consistent hash.

GossipState is responsible for maintaining the state machine of gossip protocol. In the current design, a simplified SWIM protocol without indirect ping(i.e. the suspicion mechanism) and anti-entropy are used to detect failure and maintain cluster membership.

Every node is identified with an UUID. This may cause confliction in some edge cases which we are unable to handle for now, but in most cases it is not a problem.

With the membership information, we can build a consistent hash ring to route the requests. Since Percas is a cache system, we don't need to consider consistency and data migration. Just let it be.

TODO

To integration with Percas server, there are some further works to do:

• Persist the node id and introduce incarnation mechanism to reduce inconsistency
• Maybe distinguish the commands between standalone mode and cluster mode
• Find a way to test the cluster

[tisonkun]

LGTM