Visual correspondence-based explanations improve AI robustness and human-AI team accuracy