Reducing Hallucinations in Vision-Language Models via Latent Space Steering