refactor: consolidate rectangular meshes — kernel-CDF takes the plain names

[Jammy2211]

Summary

Consolidates the rectangular mesh family down to the two meshes that matter: the kernel-density-CDF adaptive meshes (the certified gradient path, #374) and RectangularUniform (no interpolation; the HowToLens methodology illustration). The kernel implementation takes over the plain names RectangularAdaptDensity / RectangularAdaptImage, and three superseded generations leave the source: the linear empirical-CDF path (piecewise-constant likelihood in mass/shear at over-sampling 1 and on the interferometer sparse path), the spline-CDF pair (#289, shipped with a known JIT-oscillation limitation, never FD-certified), and the PCA-rotated multi-node experiment (#323, built on the spline chain — recoverable from git history if that work resumes).

Closes #402.

API Changes

The adaptive rectangular meshes keep their names but change implementation: RectangularAdaptDensity / RectangularAdaptImage are now the kernel-density-CDF meshes (formerly RectangularKernelAdapt*), gaining optional bandwidth / n_knots kwargs (defaults unchanged from the kernel classes; not free parameters — prior counts are unchanged). Results shift numerically (linear → kernel CDF; certified FoM parity 2.7e-5–6.3e-4 relative), and mass/shear gradients are now smooth in every configuration. The RectangularKernel*, RectangularSpline*, RectangularRotatedAdaptImage classes and their interpolators/geometry are removed. RectangularUniform, the InterpolatorRectangular name/module path, MeshGeometryRectangular, rectangular_edge_pixel_list_from, and all prior-config keys are unchanged — PyAutoGalaxy/PyAutoLens need zero source changes.
See full details below.

Test Plan

• pytest test_autoarray/ — 892 passed (kernel tests re-homed under the plain names; spline/rotated tests removed; RectangularAdaptImage gains its first dedicated coverage; the empty interpolator/test_rectangular.py stub replaced with the transform/areas/interpolator tests)
• PyAutoGalaxy suite against this branch — 1006 passed, zero source changes
• PyAutoLens suite against this branch — 448 passed, zero source changes
• af.Model composition: prior counts unchanged (AdaptDensity 2, AdaptImage 4, Uniform 2); bandwidth/n_knots resolve to kernel defaults at instantiation
• Grep sweep: zero references to RectangularKernel|RectangularSpline|RectangularRotated|InvertPolySpline|create_transforms_kernel|spline_deg|density_components|forward_interp in autoarray/ + test_autoarray/
• Workspace follow-up (separate PRs): update autolens_workspace_test/scripts/jax_grad variants to the plain names and re-run the strict-FD certification; stash autolens_workspace_developer spline/rotated experiments in PyAutoGut

Full API Changes (for automation & release notes)

Removed

• aa.mesh.RectangularKernelAdaptDensity / aa.mesh.RectangularKernelAdaptImage — their implementation now lives under aa.mesh.RectangularAdaptDensity / aa.mesh.RectangularAdaptImage
• aa.mesh.RectangularSplineAdaptDensity / aa.mesh.RectangularSplineAdaptImage — superseded spline-CDF meshes (Add RectangularSplineAdapt{Density,Image} meshes for gradient-based samplers #289)
• aa.mesh.RectangularRotatedAdaptImage — PCA-rotated multi-node experiment (feat: RectangularRotatedAdaptImage — PCA rotation fixes multi-source ghost-peak failure #323)
• autoarray.inversion.mesh.interpolator.rectangular_kernel (merged into interpolator.rectangular), interpolator.rectangular_spline (incl. InvertPolySpline and its JAX pytree registration), interpolator.rectangular_rotated_spline, interpolator.density_components (unwired), mesh_geometry.rectangular_rotated (MeshGeometryRectangularRotated)
• Linear empirical-CDF implementation: the old InterpolatorRectangular class body, forward_interp / forward_interp_np, the linear create_transforms and the linear transformed-grid / mappings / areas helpers
• AbstractMeshGeometry.__init__ no longer accepts spline_deg

Changed Behaviour

• RectangularAdaptDensity / RectangularAdaptImage adapt via the smooth kernel-density CDF (Enzi et al. RTU formulation) instead of the linear empirical CDF: mappings, weights, areas_transformed and edges_transformed all shift numerically (FoM parity 2.7e-5–6.3e-4 relative, certified in feat: kernel-density CDF meshes RectangularKernelAdaptDensity/Image #374); mass/shear gradients are smooth in every configuration, including pixelization over-sampling 1 and the interferometer sparse path where the linear CDF was exactly piecewise-constant
• MeshGeometryRectangular.areas_transformed / .edges_transformed always use the kernel CDF; kernel_bandwidth / kernel_knots of None (e.g. built by the uniform interpolator) fall back to the kernel defaults

Changed Signature

• RectangularAdaptDensity(shape, bandwidth=None, n_knots=None) and RectangularAdaptImage(shape, weight_power, weight_floor, bandwidth=None, n_knots=None) — None resolves to KERNEL_CDF_DEFAULT_BANDWIDTH (1.0) / KERNEL_CDF_DEFAULT_KNOTS (64); not free parameters (excluded from af.Model priors)
• InterpolatorRectangular(..., bandwidth=, n_knots=) — the class is now the kernel implementation (subclasses AbstractInterpolator directly; isinstance plot dispatch unaffected)
• interpolator.rectangular.create_transforms / adaptive_rectangular_transformed_grid_from / adaptive_rectangular_mappings_weights_via_interpolation_from — kernel-implemented, now take mesh_pixels (where applicable), bandwidth, n_knots
• adaptive_rectangular_areas_from moved from mesh_geometry.rectangular to interpolator.rectangular (kernel-implemented)

Added

• AbstractMesh subclasses expose interpolator_kwargs — extra kwargs interpolator_from forwards to interpolator_cls ({} for RectangularUniform)

Migration

• Before: aa.mesh.RectangularKernelAdaptDensity(shape=(30, 30), bandwidth=0.5)
• After: aa.mesh.RectangularAdaptDensity(shape=(30, 30), bandwidth=0.5)
• Spline / rotated meshes: removed with no in-source replacement — recover from git history (this commit's parent) if the multi-node work resumes.

Generated by the PyAutoLabs agent workflow.