Adapting Large Language Models for Parameter-Efficient Log Anomaly Detection