Fix binomial SE fallback with SVD pseudo-inverse

[Lillyakasiken]

Summary

• Fix the binomial SE fallback in irls_binomial post-processing that used sqrt(svd()$d) instead of a proper SVD-based pseudo-inverse, producing mathematically inverted standard errors when solve() fails on near-singular Fisher information matrices.
• Add testthat coverage for the SVD fallback path.

Closes #246

Test plan

• SVD fallback matches solve() on well-conditioned PSD matrices
• SVD fallback matches solve() on realistic logistic XtWX
• SVD fallback matches solve() on real abn binomial node varcov
• fitAbn returns finite positive SEs on near-singular varcov

[reinhardfurrer]

Well done!

[reinhardfurrer]

@Lillyakasiken : can you check how to pass the R_CMD_check_passed? I've approved your changes but I cannot issue this check (the others chould be checked and pass).

[Lillyakasiken]

@Lillyakasiken : can you check how to pass the R_CMD_check_passed? I've approved your changes but I cannot issue this check (the others chould be checked and pass).

The check hasn't been triggered because the CRAN::check label hasn't been applied.

[Lillyakasiken]

test-coverage: https://github.com/furrer-lab/abn/actions/runs/27133785812/job/80083378275#step:6:433

1 & 2. tests/testthat/test-build_score_cache_bayes.R:26 — buildScoreCache.bayes()
The test wraps expect_equal in expect_error to assert two buildScoreCache(..., method="bayes", ...) calls produce different outputs. Both an inner equality check and the outer "expected an error" check fired. This is an INLA-based test (line 2: skip_on_cran()) — it doesn't run in the R_CMD_check matrix jobs.

3. tests/testthat/test-tographviz.R:181 — duplicate snapshot file name
Error: Snapshot file names must be unique. "graph.dot" has already been used.
Looking at the test file, the name "graph.dot" is used at lines 171, 181, and 191. Pre-existing test bug — should be graph2.dot, graph3.dot.