SafeMind: Benchmarking and Mitigating Safety Risks in Embodied LLM Agents