feat(xpu): support vllm_kunlun backend

[loveleaves]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Add support for the vllm_kunlun plugin (Kunlunxin XPU backend for vLLM), following the existing vllm_ascend / vllm_metax pattern:

• swift/utils/import_utils.py: add is_vllm_kunlun_available().
• swift/utils/__init__.py: export it (both TYPE_CHECKING and lazy import map).
• swift/pipelines/infer/rollout.py and swift/rlhf_trainers/vllm_client.py: when vllm_kunlun is installed, use PyKunlunCommunicator as PyNcclCommunicator for weight-sync communication.

No behavior change on other backends: the new import path is only taken when vllm_kunlun is importable.

Experiment results

N/A (import-level device backend adaptation).

[tastelikefeet]

Hi, 问下是 kunlun 的官方同学吗？是否方便支持 dockerfile 呢：
https://github.com/modelscope/modelscope/tree/master/docker
这样更方便开发者使用