Similar to how we did for SERVER-100762, we should transfer the OperationMemoryUsageTracker to the ClientCursor when a sharded query produces more than a single batch, and write a test that shows this works.

Do this for aggregate() to start, we may need to add more tickets for find(), distinct().