Inducing anxiety in large language models increases exploration and bias