Supervised Fine-Tuning or Contrastive Learning? Towards Better Multimodal LLM Reranking