Regression (b88): OOM on parallel limit operation

[Brian Goetz]

> There are definitely things we need to tweak here.

Tweaking update: this led us to discover a significant memory leak -- 
thanks for calling it to our attention.  We've got a patch (not yet 
pushed, still working on it) that brings the OOME threshold for your 
example program down from a heapsize of 450M down to 9M so far, but 
we're not done yet.

I still think 9M is way too high for solving this problem, even given 
complete buffering and pathological splitting.  So I'll continue to beat 
on it.  Still, all the things I said this morning still hold -- 
sequential generators and parallel limit are fundamentally a bad 
combination.