Transformer Language Models without Positional Encodings Still Learn Positional Information