Learning Robust Output Control Barrier Functions from Safe Expert Demonstrations