Performance improvements to shuffle operations: in particular, preserve

RDD partitioning in more cases where it's possible, and use iterators
instead of materializing collections when doing joins.