vLLM adds day-zero reinforcement learning support with Megatron

Day-0 RL support, too: vLLM + Megatron.

@inferact’s work covers training–inference consistency for Kimi Delta Attention, alongside the cache and state management needed to serve long-running agents.

From serving the model to continuing its training, the open stack is ready for the next step.