Towards Eliminating Hard Label Constraints in Gradient Inversion Attacks