A gradient flow approach to decentralized output feedback optimal control