NLPositionality: Characterizing Design Biases of Datasets and Models