Efficient Inference of Average Treatment Effects in High Dimensions via Approximate Residual Balancing