Revisiting Softmax Masking: Stop Gradient for Enhancing Stability in Replay-based Continual Learning