Optimistic PAC Reinforcement Learning: the Instance-Dependent View