Unveiling Attractor Cycles in Large Language Models: A Dynamical Systems View of Successive Paraphrasing