Index consistency check at end of sharding tests isn't robust to ShardNotFound

The balancing_sessions_collection.js intentionally removes one of the shards before shutting down the ShardingTest. It is possible for mongos to refresh its ShardRegistry after choosing to target all shards with the $indexStats aggregation and then triggers a ShardNotFound error.

One proposal to address this issue is to retry running the $indexStats aggregation on ShardNotFound errors.