Support stronger durablility guarantee

Sometimes we need to ensure that the data is not lost in the scenario where 2 replicates cannot be recovered at the same time(we have 3 replicates). This requires 3 replicates to be committed before response to client.

But currently, as long as there are 2 replicates to be committed, the write rpc success is returned to the client. Maybe we can add a gflag to indicate whether all replicates need to be committed before applying(response to client write success).