Read Between the Layers: Leveraging Intra-Layer Representations for Rehearsal-Free Continual Learning with Pre-Trained Models