Provably Efficient Risk-Sensitive Reinforcement Learning: Iterated CVaR and Worst Path