LLMSteer: Improving Long-Context LLM Inference by Steering Attention on Reused Contexts