Policy Disruption in Reinforcement Learning:Adversarial Attack with Large Language Models and Critical State Identification