Relax tolerance in Qwen Scope loader test to fix flaky CI

[robbiebusinessacc]

Description

test_qwen_scope_sae_huggingface_loader_with_mocked_download intermittently
fails in CI (e.g. the two runs linked in #688). It compares two mathematically
equal matmuls: the loader stores W_enc/W_dec as .T.contiguous(), while the
reference in the test multiplies a transposed view (raw_W_enc.T). Over the
2048- and 32768-dim reductions, a contiguous tensor and a transposed view can
accumulate float32 in a different order depending on the BLAS backend, so the
two results agree only up to rounding. The default assert_close tolerance
(rtol=1e-5, atol=1e-8) is too tight to absorb that, which is what danra
diagnosed on the issue.

This compares with rtol=1e-3, atol=1e-2, matching the existing precedent for a
matmul comparison elsewhere in this file. The tolerance is generous enough to
absorb the accumulation noise but still catches a real loader regression, which
would be orders of magnitude off (verified locally: a 1% weight-scale error and
a wrong-bias error are both still rejected).

Note: the failure is BLAS/threading-dependent and reproduces on CI but not on
local arm64 macOS (where the two layouts produce bit-identical results), so this
is verified by the linked CI logs plus the regression check above rather than a
local repro.

Fixes #688

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 tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

You have tested formatting, typing and tests

• I have run the formatter and linter (ruff format + ruff check) on the change.

[chanind]

Looks good, thanks for this!