Fix index reincarnation routing bug

[nadav-govari]

Description

Node based routing introduced a bug - there was no index_uid tracking, and a new index_uid didnt clear old routing entries. As a result, deleting an index and recreating it with the same name, which is normally fine, would result in unavailability.

This fixes that by:

1. clearing the routing entries for an index on a greater index incarnation
2. Broadcasting 0 shards open on that node.
3. deleting the key from chitchat after that, which removes it from the routing tables
4. Eventually, once chitchat converges, 0 open nodes for the index
5. If a new request comes in for that index, it should see the old incarnation of the index, but with 0 nodes, and call the control plane, which triggers a merge from shards and gives us what we want

Added a unit test and an integ test

[guilload]

One cycle is too short. Let's just use set_self_key_value_delete_after_ttl and get rid of the pending_removal logic.

[nadav-govari]

Much cleaner

[guilload]

Logic is sound. Code is inelegant. The old routing table code also had useful comments:

match self.index_uid.cmp(index_uid) {
            // If we receive an update for a new incarnation of the index, then we clear the entry
            // and insert all the shards.
            std::cmp::Ordering::Less => {
                self.index_uid = index_uid.clone();
                self.clear_shards();
            }
            // If we receive an update for a previous incarnation of the index, then we ignore it.
            std::cmp::Ordering::Greater => {
                return;
            }
            std::cmp::Ordering::Equal => {}
        };