[NO MRG] Try Rick's example in the benchmark

[jakirkham]

Note: This is merely intended to be illustrative. No need for merging

This tries @rjzamora's example ( dask/dask#7615 (comment) ) as shown in the diff below. So is a follow up to the profile in issue ( #137 ).

To make this more comparable to existing work with shuffle, this makes a couple of notable tweaks to Rick's example. First it uses the same input data that is used with shuffle. However this doesn't show much in a single (or even a few) iterations. So this does 20 iterations instead of the typical 5 iterations with shuffle.

With these modifications, this starts to look more analogous to the results that we see in a normal shuffle like in issue ( #140 ) (even though this is not a shuffle). Namely time is spent mostly in communication followed by finished task transitions. IOW backing up what was found before ( #139 (comment) ).

cc @quasiben