On the Low-Complexity of Fair Learning for Combinatorial Multi-Armed Bandit