Guarantee GaussianKernel regularization matrix is PD

[Jammy2211]

Summary

• Symmetrise and add trace-scaled diagonal jitter to the inverted covariance matrix in GaussianKernel.regularization_matrix_from so downstream cholesky(regularization_matrix_reduced) in log_det_regularization_matrix_term cannot fail on floating-point noise.
• Root cause: when the Gaussian scale is large relative to pixel spacing, cov = exp(-d²/(2·scale²)) + 1e-8·I is numerically near rank-1. xp.linalg.inv(cov) then returns a matrix that is mathematically PD but in floating point is slightly asymmetric with tiny negative eigenvalues, crashing the cholesky in FitDataset.log_evidence. This reliably reproduced under PYAUTOFIT_TEST_MODE=2 for 3 autogalaxy pixelization scripts with LogUniformPrior(1e-6, 1e6) on scale.
• Uses the same xp dispatching as the rest of the module, so the fix works for both the numpy and JAX backends without branching.

Test plan

• Repro the pre-fix LinAlgError with PYAUTOFIT_TEST_MODE=2 + wiped output dir
• Post-fix: 3 previously failing scripts (autogalaxy imaging/features/pixelization/modeling.py, interferometer/features/pixelization/modeling.py, howtogalaxy/chapter_4_pixelizations/tutorial_5_model_fit.py) all exit 0
• No-regression: autolens pixelization imaging/features/pixelization/modeling.py still passes
• pytest test_autoarray/inversion/regularizations/ — 33 passed (including test_gaussian_kernel.py)

Paired with the SLOW-skip PRs on autogalaxy_workspace and autolens_workspace under the same branch name.

Generated with Claude Code