Mid-Level Visual Representations Improve Generalization and Sample Efficiency for Learning Active Tasks