Sample scenario for repro

In a n-sharded cluster, an empty collection targeted with a createCollection({key: hashed}) should produce a routing table with the following chunk distribution:

• if gOneChunkPerShardEmptyCollectionWithHashedShardKey is enabled (>= 7.2), n chunks and 1 chunk per shard
• if gOneChunkPerShardEmptyCollectionWithHashedShardKey is not enabled (<7.2), 2*n chunks and 2 chunks per shard

The logic is nevertheless flawed: the feature flag is queried multiple times during the generation of the routing table - and its value may be unstable across reads. As a consequence of this, we might end up in a state where shardCollection generates n chunks (feature flag on) but only distributes them across n/2 shards (feature flag off) when the DDL is executed concurrently with a FCV downgrade.