fix: test-mode-gate singular/non-PD inversion crash (release tail)

[Jammy2211]

Closes part of PyAutoArray#388.

Summary

Return a benign dummy from the inversion reconstruction / log-det sites only when autoconf.is_test_mode() is active, instead of raising on a singular/non-PD curvature-regularization matrix. In test mode the fitted model is fabricated / under-converged and the inversion output is discarded, so an unguarded test-mode path (search chaining, preloads, result construction) should not crash — while the real likelihood path already resamples such solutions via FitException.

Normal operation is byte-for-byte unchanged: all four sites re-raise exactly as before (LinAlgError / InversionException), so real-mode numerics and search resample behaviour are untouched. numpy-only by construction (the JAX linalg path returns NaN rather than raising). No conditioning floor — nothing perturbs real evidences/reconstructions.

Sites gated on is_test_mode():

• inversion_util.reconstruction_positive_negative_from (xp.linalg.solve)
• inversion_util.reconstruction_positive_only_from (fnnls except)
• abstract.log_det_curvature_reg_matrix_term (cholesky)
• abstract.log_det_regularization_matrix_term (cholesky)

Corrective-PR exception (Heart RED)

Opened under the human-authorized corrective-PR exception (AUTONOMY.md), authorized in-session 2026-07-14.

• Named RED reason: release validation FAILED (stage integrate)
• Causal map: integrate stage FAILed on interferometer/.../multi_gaussian_expansion/slam.py + imaging/features/pixelization/cpu_fast_modeling.py (run 29341418859) → both crash on an unguarded singular inversion in test-mode chained paths → 4-site is_test_mode() dummy guard → this PR. Follow-up to PyAutoLens#607, which guarded the real/likelihood path but not these test-mode paths.
• Scope: this one pending-release PR only; no merge, issue close, release, or rehearsal in this action. Merge is a separate human act.
• Sibling RED reason remains: PyAutoArray: on branch feature/ticks-minus-in-math (not main) — Heart stays RED until that is also resolved and integrate validation re-runs.

Caveat on causal strength

Unlike #607 (deterministic repro), these two FAILs did not reproduce on clean main (TEST_MODE=1 and =2 both pass; the crash is flaky/non-deterministic). This is a test-mode-gated dummy, not a real-path correctness change. It plausibly removes the FAILs; confirmation comes only from the post-merge fresh-wheels re-validation of the integrate stage.

Validation checklist

• New numpy-only unit tests: singular matrix raises normally; under monkeypatched PYAUTO_TEST_MODE returns finite dummies.
• test_autoarray/inversion/inversion/ (util + abstract): 41 passed.
• Full test_autoarray/inversion/: 233 passed.
• No JAX in unit tests (per repo rule); JAX path covered by *_workspace_test parity scripts.
• Post-merge: fresh wheels + re-run release integration on main + new Heart verdict (required before any release decision).

Self-review: CLEAN — commit 48fae1ac, 3 files (+101/−12), no mixed scope.