Are LLMs Failing to Match with Suffix in Fill-in-the-Middle (FIM) Code Completion? Horizon-Length Prediction: A New AI Training Task to Advance FIM by Teaching LLMs to Plan Ahead over Arbitrarily Long Horizons