Deduplicating Training Data makes Language Models Better