The Justice Department wrote that “the creative possibilities and public benefits” of training large language models on such content “far outweigh any competitive harm”