No GPU left behind: Unlocking Efficiency with Co-located vLLM in TRL