[Tests][Eagle3] Extend vLLM test cases with conversion step

[dsikka]

Summary

• Updates vLLM token assertion
• Updates how vllm is detected, if installed
• Extends the vLLM test flow to add an additional test with conversion
• Converts a speculators model and a model from the Eagle3 repo
• Runs the converted models in vllm and asserts for valid tokens

[github-actions]

📦 Build Artifacts Available
The build artifacts (`.whl` and `.tar.gz`) have been successfully generated and are available for download: https://github.com/vllm-project/speculators/actions/runs/17244541644/artifacts/3855588296.
They will be retained for up to 30 days.
Commit: d99db49

[shanjiaz]

🎉

[markurtz]

One major issue that I left as a comment on the code, otherwise there are several minor cleanups that we could get in for this or for a later PR:

• The sampling params use temperature and top_p which causes extra, unneeded variability in the tests
• The test is asserting that the number of returned tokens has to be 20. Should be updated to >1 and <=20 or we should null out the stop tokens so it's forced to go to 20
• There's no cleanup of the vLLM resources (del llm, gc.collect, empty_cache, etc). A parameterized fixture with a yield would ensure both setup and teardown always run
• Currently only checking for vLLM imports are available, we likely would want to put an extra case on there for torch.cuda / GPU availability so we don't waist resources
• We could require in the dev environments hf_transfer and set HF_HUB_ENABLE_HF_TRANSFER for faster downloads
• _run_vllm_engine runs as a helper, so any failures within there are not going to surface properly with pytest
• Enabling a test config/settings longer term would be good so we can pull down things like gpu_memory_utilization and have a central place to update them for workers running tests
• Base comparisons / interface comparisons for what the output from conversion should look like before running in vLLM

[markurtz]

Missed one of the pytest skips in the original review, approving with a note towards the potential future improvements we should get onto the backlog