Instead we tolerate racing to cache sub-arrays, like we do for individual values.
This does mean two or more threads might create the same sub-array, with only one
winning, but this should be rare with later calls picking up the same sub-array.