A Capabilities Approach to Studying Bias and Harm in Language Technologies