[DONT MERGE] Branch for vllm nightly ci

[Jackmin801]

Note

Introduces token-in chat completions and adapts to vLLM nightly module/layout changes.

• Adds OpenAIServingChatWithTokens and /v1/chat/completions/tokens route; allows overriding prompt_token_ids from request tokens and supports streaming/full responses
• Updates imports to new vLLM OpenAI modules (chat_completion.protocol/serving, engine.protocol) and adjusts server wiring
• Monkeypatches:
  • PrometheusStatLogger init to bypass DP-mode LoRA check while restoring LoRA metrics
  • OpenAIServingModels.load_lora_adapter to reuse/update existing adapters by name
  • LRUCacheWorkerLoRAManager to avoid redundant per-request loads and manage cache activation/eviction
• Customizes init_app_state and worker proc to ensure patches are applied in multi-API-server setups
• Test utils now only consider SUCCESS step lines matching the reward pattern for numeric checks

^{Written by Cursor Bugbot for commit b16f34a. This will update automatically on new commits. Configure here.}