Scrutinizing XAI using linear ground-truth data with suppressor variables