Perplexity from PLM Is Unreliable for Evaluating Text Quality