Breaking the Length Barrier: LLM-Enhanced CTR Prediction in Long Textual User Behaviors