Remove deprecated prompt_token_ids wrapping in vLLM backend

[sihyeonn]

Context

Follow-up to #1173 which upgraded vLLM to >=0.11.0. A few call sites still used the old prompt_token_ids wrapping patterns that are no longer needed.

Relates to #1002

What changed

vLLM's PromptType includes list[int] directly in DecoderOnlyPrompt, so we can pass token ID lists to LLM.generate() without any wrapping:

PromptType = DecoderOnlyPrompt | EncoderDecoderPrompt
DecoderOnlyPrompt = str | TextPrompt | list[int] | TokensPrompt | ExplicitEncoderDecoderPrompt

Removed:

• {"prompt_token_ids": ...} dict wrapping in the ray data-parallel path (vllm_model.py)
• TokenInputs(prompt_token_ids=...) wrapping + its import in the single-GPU path (vllm_model.py)
• {"prompt_token_ids": ...} dict wrapping in the LLM-as-judge vLLM path (llm_as_judge.py)

Now all three call sites just pass list[list[int]] directly, which is simpler and works across vLLM 0.11 through 0.17+.

Test plan

• Verified against vLLM docs that list[int] is a valid PromptType from 0.10.2 onwards
• Existing vLLM tests in the repo should continue to pass (no behavioral change, just removing unnecessary wrapping)