LATTE: Label-efficient Incident Phenotyping from Longitudinal Electronic Health Records