GridFS should be able to shard fs.chunks collection using hash base sharding over file_id field.
This would make it much more efficient in terms of even distribution of data across nodes.
Keep the uniqueness of pair file_id + n but be able to shard based on the hash.