Fints: Efficient Inference-Time Personalization for LLMs with Fine-Grained Instance-Tailored Steering