Assessing Risks of Large Language Models in Mental Health Support: A Framework for Automated Clinical AI Red Teaming