Group and Unique use an arbitrary assumption that there will be about
10% as many groups as input tuples --- perhaps someday we can refine this.