Reducing Hallucinations in Large Vision-Language Models Via Latent Space Steering