Towards Parameter-Efficient Integration of Pre-Trained Language Models In Temporal Video Grounding