Fix QueryMaintenance calls in Aggregators

A recent change, IMPALA-110 (part 2), refactored PartitionedAggregationNode into several classes, including GroupingAggregator. During this refactor, code that makes local allocations while evaluating exprs was moved from the ExecNode (now AggregationNode/StreamingAggregationNode) into the new type Aggregator, but code related to cleaning these allocations up (ie QueryMaintenance()) was not, resulting in some queries using an excessive amount of memory.