Demonstrate an adaptive hierarchical min reduction

[frankmcsherry]

This PR demonstrates through examples/min.rs a hierarchical aggregation that adapts to the group size. For an input collection of (key, val) records, it produces (key, val, lvl) results with the properties that:

1. The value of lvl is one of two values, those which sandwich the logarithm of the count for the key (under your favorite base).
2. Each update to the input results in at most a bounded number of updates to the output, generally by at most the base used above for the logarithm (e.g. that many values may get bumped to a higher level, or drawn back down to a lower level).

The algorithm is based on the idea that one can think of putting values into bins at each of various levels, defined by (key, lvl, hash(val) % base.pow(lvl)). We'll use (key, count) to mark various bins as "full", and ask each (key, val) to iteratively discover which level they should be at (starting at zero). With some care in each of these steps, they seem to have the property (modulo bugs, fundamental errors) that updates to the input result in bounded work to update the output.

cc: @petrosagg @ggevay