Backpropagation through the Void: Optimizing control variates for black-box gradient estimation