Retry SELECT after placement metadata change

SELECT commands may error out during operations such as a rebalance and `master_update_node` because they do not take any shard metadata locks. Taking locks is mostly undesirable because it might cause a long-runnning SELECT to block a failover. An alternative is to retry the SELECT with fresh placement metadata.

We already have the necessary logic to retry SELECTs. We mainly need to reload the placement metadata and retry with the new placement metadata if it changed.

Potentially, we could avoid SELECTs from erroring out completely during rebalance and switchover.