Reliable or Deceptive? Investigating Gated Features for Smooth Visual Explanations in CNNs