[RFC] Reward model (value-head RM) support in infer/deploy via vLLM

Hi, I'd like to pick up this item from the #5721 roadmap: vLLM/SGLang support for classification models / RM.

Current state on main:

• VllmEngine already maps embedding / seq_cls / reranker to vLLM pooling tasks via encode_task_mapping, so classification models are covered — but there is no task type for reward models.
• swift infer / swift deploy can't load a value-head RM directly, and sglang_engine.py only special-cases embedding.
• The only ways to score an RM today are the in-process transformers forward in DefaultRMPlugin (swift/rewards/rm_plugin.py), or deploying it separately with a bare vllm serve.

Proposal:

1. Add a reward task type end-to-end: map it to the vLLM pooling runner in VllmEngine and return pooler scores from infer.
2. Expose it in the infer/deploy CLIs so swift deploy --task_type reward works out of the box.
3. Optionally, let GRPO's reward plugin call this engine, so the RM can run on separate GPUs with much higher throughput during training.
4. SGLang support as a follow-up once the vLLM path is settled.

I'd like to work on this. Before I start — does this scope look right, and would reward be the preferred task_type name? Thanks!

Start by reading VllmEngine's encode_task_mapping and the existing pooling-task handling, then inspect swift infer/deploy task-type definitions and sglang_engine.py. Compare these paths with swift/rewards/rm_plugin.py to define the smallest vLLM-only scope. Done should mean the agreed reward task works through infer and deploy, with CLI behavior and tests covering the supported path; SGLang and GRPO integration are explicitly resolved as in- or out-of-scope.