Risks from Language Models for Automated Mental Healthcare: Ethics and Structure for Implementation