Towards Low-Resource Harmful Meme Detection with LMM Agents