Towards Fundamental Limits of Multi-armed Bandits with Random Walk Feedback