Currently all of the accumulator initialization is performed as part of the accumulation itself. This is not applicable for $topN/$bottomN, because they can take an "n" parameter which varies between groups. We want to extend the hashagg stage so that it can initialize a group based on an expression.