On the Risk of Evidence Pollution for Malicious Social Text Detection in the Era of LLMs