Measuring Social Biases in Grounded Vision and Language Embeddings