Bridging Distributional and Risk-sensitive Reinforcement Learning with Provable Regret Bounds