Restore tokenizer special-id patches in attribution tests

[Andrey170170]

Summary

This PR prevents TransformerLens attribution tests from leaking tokenizer class state across tests.
The Gemma and Llama dummy-model tests temporarily patch tokenizer.all_special_ids so attribution treats token 0 as special. Previously, the tests assigned directly to the tokenizer class property without restoring the original value.
This change:

• adds a small patch_tokenizer_special_ids(...) context manager
• removes permanent all_special_ids mutation from dummy model loaders
• wraps the relevant Gemma/Llama dummy attribution tests so the original tokenizer class property is restored after each test
• reuses the helper in the Llama tests via the existing Gemma test imports

Why

This avoids leaking global tokenizer class state across tests and follows the safer restore pattern already used by the NNSight attribution tests.

Tests

Run:
python -m ruff check tests/test_attributions_gemma.py tests/test_attributions_llama.py
python -m ruff format --check tests/test_attributions_gemma.py tests/test_attributions_llama.py
python -m pytest tests/test_attributions_gemma.py::test_small_gemma_model tests/test_attributions_llama.py::test_small_llama_model

Results:

ruff check: passed
ruff format --check: passed
pytest: 2 passed

Not run:

• full pytest suite
• full pyright

[Andrey170170]

CI note: the failing build check appears unrelated to this PR.

The failure is in the check types step (python -m pyright). Current main is failing the same step as well:
https://github.com/decoderesearch/circuit-tracer/actions/runs/26090319535

I reproduced locally that latest pyright (1.1.409) reports many torch export errors, while the currently declared lower-bound version passes:

• uv run python -m pyright → fails with reportPrivateImportUsage on torch symbols
• uv run --with pyright==1.1.403 python -m pyright → passes

Likely cause is the unpinned dev dependency pyright>=1.1.403 resolving to a newer pyright version.

[hannamw]

Yes, that's the conclusion I reached as well - the issue seems to have appeared in a pyright past 1.1.406. The torch issues appear to be noise, and can been suppressed, but some of the issues are legitimate. I've kept pyright unpinned, but suppressed the torch issues, while fixing the other ones, in #99 . I think it should be possible to merge this now!

[hannamw]

This looks good, and seems like a sensible change. Thanks for contributing!