Learning to Bid Optimally and Efficiently in Adversarial First-price Auctions