MM-Ego: Towards Building Egocentric Multimodal LLMs