SoK: Towards Security and Safety of Edge AI