As of SERVER-23654's resolution, the $facet stage currently puts everything into one giant buffer before passing results through the sub-pipelines. We should allow the sub-pipelines to process one batch at a time, removing the constraints of fitting the result set into the memory limit of 100MB.