etcdserver: commit before sending snapshot

[xiang90]

Fix #5857.

Reproduce:

1. write some v2 keys to force a snapshot
2. remove a member
3. add the member back
4. kill the added member immediately before a new snapshot is triggered or consistent index is ever updated.

Then the newly added member will not be able to restart anymore.

This pr fixes it.

[xiang90]

/cc @binwiederhier @heyitsanthony

[heyitsanthony]

lgtm
is it necessary to do a full Commit() to persist down to disk or is it only so that saveIndex is triggered so Snapshot() has the most recent index?

[binwiederhier]

Can confirm that it worked for me.

[gyuho]

lgtm.

[xiang90]

is it necessary to do a full Commit() to persist down to disk or is it only so that saveIndex is triggered so Snapshot() has the most recent index?

Both. We need to send a snapshot contains the full synced data.

But in this case, the commit() is just for the metadata since when we call db.Snapshot(), it will do the db.sync() for us, however without updating the in-memory metadata like consistentIndex.