vLLM upgrade required matching token probability calculations

Original: vLLM V0 to V1: Correctness Before Corrections in RL

Source: huggingface.co ↗