Inference-Aware Fine-Tuning for Best-of-N Sampling in Large Language Models