Large arrays take a long time to copy during grafting

[azf20]

We really need to do something about large arrays - there are some grafts in shard_i involving subgraphs with large arrays where a batch to copy data took over an hour when each batch should really only take 5 minutes. This PR tries to mitigate that, but a lot of the data we store comes from large arrays, and we need to do something better there.