Loosen float32 tolerance in edge-effects mutation test

[YHC66]

Description

tests/unit/tools/test_attribution_patching.py::test_edge_effects_mutation_only_changes_the_perturbed_writers_edges fails with torch 2.14.1 (the latest release; pyproject.toml allows torch>=2.6, the lock pins 2.11.0):

>   assert mutated_scores[edge] == pytest.approx(expected)
E   assert 0.000663460697978735 == 0.00066345982...8213 ± 6.6e-10

The expected value is baseline_scores[edge] + shift, and for this edge the two partly cancel (≈ -5.3e-4 + 1.19e-3), so float32 rounding error relative to the result is amplified. The worst relative error over all edges is 1.3e-6 with torch 2.14.1 and already 5.3e-7 with torch 2.11.0, against pytest.approx's default rel=1e-6. The library code is fine; only the tolerance is too tight for float32.

This switches the assertion to abs=1e-6, the tolerance the other edge-score comparisons in the same file already use. The shifts in this test are around 1e-3, so the check still catches a mutation that is missing or applied to the wrong edge.

Observed on macOS arm64 (CPU). With torch 2.14.1 + transformers 5.17 this was the only failure in tests/unit on dev (6affbe99): 6933 passed, 1 failed.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• My changes generate no new warnings
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[jlarson4]

Looks good @YHC66, thanks for this slight tolerance adjustment