Pull VllmConfig construction up to Policy

[Jack-Khuu]

Move VllmConfig construction into EngineConfig at the Policy level instead of deferring to the PolicyWorker.
This simplifies the PolicyWorker construction logic from:

class PolicyWorker(ForgeActor):
    vllm_args: EngineConfig | Mapping = EngineConfig()
    ...

    + def __post_init__() ...

to just

class PolicyWorker(ForgeActor):
    vllm_config: VllmConfig

---

Misc:

• While doing this it also cleans up the ambiguous intermediate vllm_args from Policy
• Fixed SamplingConfig bug where temp and top_p weren't being passed down from yaml cc: @DNXie

---

python -m apps.grpo.main --config apps/grpo/qwen3_1_7b.yaml
python -m pytest tests/unit_tests/test_policy_config.py
python -m apps.vllm.main --config apps/vllm/llama3_8b.yaml

[joecummings]

One comment, but looks clean!

(Sorry about the merge conflicts :) )

[allenwang28]

I did something like this in #151, wondering what you think about that approach?

[allenwang28]

actually i think i like yours better

[Jack-Khuu]

@allenwang28 Ah I missed this change in #151, part of reason for this mini PR was actually your temp work for testing vllm

I considered doing it as a classmethod like you have it at first, but EngineConfig.as_engine_args(self.engine_config) felt a lil verbose

[DNXie]

Thanks for spotting and fixing the issue with SamplingConfig! The code LGTM