ViSaRL: Saliency-Guided Visual Representations for Robot Learning