TokenAlign: Token-Level Inference-Time Guidance for Reducing Hallucinations in Vision-Language Models