On Background Bias of Post-Hoc Concept Embeddings in Computer Vision DNNs