This change will help the pd profile handler to be able to make decision about disaggregation based on the precise prefix scorer's logic.
Tokenization will be done in the scorer's code before calling the GetPodScores function.
Relevant issue is in llm-d-inference-scheduler project.