Make the Pertinent Salient: Task-Relevant Reconstruction for Visual Control with Distractions