It's too sensitive to outliers, but we may as well add it to the branch anyway. This way we centralize the different strategies tried.

There may be some way to make this idea work, but nothing is coming to mind at the moment.

Included in this diff is the prototype code for collecting streamed histograms. This could be useful in getting a more detailed look at the distributions as we start to vary the parameters to the query shapes.

We can also add just a plain list of all the latencies, though for queries with many runs, the histograms provide a more intuitive first glance.