[skyrl-train][algorithm] Apply inference engine top-k mask to trainer

[erictang000]

As described in the Deepseek v3.2 Tech report - if using top_p or top_k sampling, the sampling mask should be propagated to the trainer to avoid drifting off policy in the trainer compared to the inference engine.

[devpatelio]

@erictang000 From my understanding, there is no actual way to extract the mask or logits directly from vLLM (can't even recompute the mask even if you wanted to) - I'll implement a patch for now but will also open an issue in vLLM so we can remove the patch in the future.