This is a custom partitioner (used in conjunction with
AssignIndexFunction to repartition a RDD.
Unlike a standard .repartition() call (which assigns partitions like [2,3,4,1,2,3,4,1,2,...] for 4 partitions],
this function attempts to keep contiguous elements (i.e., those elements originally in the same partition) together
much more frequently. Furthermore, it is less prone to producing larger or smaller than expected partitions, as
it is entirely deterministic, whereas .repartition() has a degree of randomness (i.e., start index) which can result in
a large degree of variance when the number of elements in the original partitions is small (as is the case generally in DL4J)