Add unit tests for VLLMModel sampling params and config wiring

[jayzuccarelli]

Note: This resubmits #1276 (same commit). I closed it temporarily while confirming my employer's open-source contribution policy — that's now cleared.

What

Adds unit tests for VLLMModel, as suggested in #724. These cover the pure-Python logic that doesn't need a GPU or a real vllm install (the existing @requires("vllm") gate still applies, so they run under the dev-gpu test env like the other cases in this file):

• _generate → SamplingParams mapping — this is the boundary with vllm that Fix revision arg for vLLM tokenizer #721 showed can break silently. Tests pin down:
  • the greedy scoring path (generate=False): temperature=0, max_tokens=1, prompt_logprobs=1, detokenize=False;
  • the generation path: n, max_tokens, stop, and logprobs are mapped from the runtime args, with temperature carried over from generation_parameters;
  • the num_samples > 1 + temperature == 0 guard raises ValueError.
• __init__ config wiring — add_special_tokens, max_length, and the tokenizer property are taken from the config (construction kept offline by patching out model/tokenizer creation).

Notes

• No new dependencies; mirrors the existing mock/patch style in this file. A small _RecordingSamplingParams stand-in is used so the asserted values are actually checked rather than swallowed by a bare Mock.
• pytest tests/unit/models/vllm/test_vllm_model.py → all green; ruff check / ruff format clean.

Closes #724