Steering LVLMs via Sparse Autoencoder for Hallucination Mitigation