DropEdge not Foolproof: Effective Augmentation Method for Signed Graph Neural Networks