Improving Diversity of Demographic Representation in Large Language Models via Collective-Critiques and Self-Voting