Same model, better performance: the impact of shuffling on DNA Language Models benchmarking