Learning Generalizable Agents via Saliency-Guided Features Decorrelation