Refined Analysis of FPL for Adversarial Markov Decision Processes