Adaptive Prompting for Continual Relation Extraction: A Within-Task Variance Perspective