Ground Truth Evaluation of Neural Network Explanations with CLEVR-XAI