On the Difficulty of Token-Level Modeling of Dysfluency and Fluency Shaping Artifacts