Efficiently Learning and Sampling Interventional Distributions from Observations