Only load Verifier model if attachment_mode is 'full'

[fynnsu]

Currently, if verifier_attachment_mode is "full" or "train_only" we load the full verifier model on init.

However, for the "train_only" case, we want to only load the parts of the model that are needed. The logic to do so, is meant to reside in the subclasses of SpeculatorModel.

This change, makes it so that we don't auto-load the full verifier model if "train_only" is selected.
Additional changes:

• SpeculatorModel.attach_verifier() no longer returns the loaded verifier (since it might not exist)
• EagleSpeculator.attach_verifier() directly calls resolve_verifier() to load the verifier model and extract its components.
• test_speculator_model_attach_verifier_invalid needed to be updated to reset the model between subtests because otherwise the verifier_attachment_mode gets set and kept by the previous try-excepts blocks.

[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/18918018372/artifacts/4409360027.
They will be retained for up to 30 days.
Commit: 2b17856

[shanjiaz]

cool

[brian-dellabetta]

bummer that we have to have all those # type: ignore[assignment,attr-defined] comments

[fynnsu]

bummer that we have to have all those # type: ignore[assignment,attr-defined] comments

Yeah I'm honestly not too sure what the deal with those is. We might want to try removing them at some point, but it's also possible this eagle3 file will be removed in favor of the new training one