vllm - 💡(How to fix) Fix prompt_tokens_details.cached_tokens always reports prompt_tokens - 1 in disaggregated prefill/decode mode

Your output of `python collect_env.py` here

---

## Summary
                                                                                                     
  When running vLLM in disaggregated prefill/decode mode with `--enable-prompt-tokens-details`, the  
  `usage.prompt_tokens_details.cached_tokens` value is always reported as `prompt_tokens - 1`,       
  regardless of whether any prefix cache hit actually occurred.                                      
                  
  This happens even on the very first request with a unique prompt that cannot possibly have a prefix
   cache hit.
                                                                                                     
  ## Expected behavior

  `cached_tokens` should reflect actual **prefix cache hits from prior requests** — matching the     
  OpenAI API semantics for this field. For a cold request with no prior matching prefix,
  `cached_tokens` should be `0` (or at most a small constant if the cache contains a system-prompt   
  prefix).