[Jammy2211]

Phase 1 of the @rhayes777 API-audit campaign (epic #415, task #416). Fixes the
ConstantSplit × RectangularUniform crash reported in #332.

What was wrong

Rectangular meshes do not support split regularization — their interpolators provide no
split-cross mappings. The workspace prose already said so
(autolens_workspace/.../slam.py:18: "AdaptSplit / ConstantSplit regularization is only
valid on irregular meshes such as Delaunay"), but nothing enforced it, so the combination
failed deep inside the inversion in two different ways:

Mesh	Interpolator	Before
RectangularUniform	InterpolatorRectangularUniform	AttributeError: ... has no attribute '_mappings_sizes_weights_split'
RectangularAdaptDensity	InterpolatorRectangular	IndexError: index 4 is out of bounds for axis 0 with size 4
RectangularAdaptImage	InterpolatorRectangular (inherits)	same IndexError

Nine combinations were affected (3 rectangular meshes × 3 split schemes —
ConstantSplit, AdaptSplit, AdaptSplitZeroth), not the one reported.

What changed

Pixelization.__init__ rejects the pairing, naming both the mesh and the regularization and
pointing at what does work. Driven by two capability flags rather than isinstance chains, so
a new mesh declares its own capability:

• AbstractMesh.supports_split_regularization (default True; False on RectangularAdaptDensity,
  inherited by the other two)
• AbstractRegularization.is_split_regularization (default False; True on the three split schemes)

Also removes InterpolatorRectangular._mappings_sizes_weights_split. It returned the plain
4-corner mappings, commented as "split regularization reuses the same mappings" — which is not
true. reg_split_from expects the split-cross structure, and that pass-through is what produced
the IndexError one frame later. Leaving it would invite someone to mirror it onto the uniform
class.

Why this exception type

autoarray.exc.PixelizationException is deliberately not a FitException.
autogalaxy.exc.PixelizationException is one, and fitness.py:256 converts those into
resample-rejects — so raising the wrong class would make a search silently discard every sample
instead of reporting the misconfiguration. A test pins this.

API Changes

Behavioural change, no signature change.

• Pixelization(mesh=<rectangular>, regularization=<split>) now raises
  PixelizationException at construction instead of failing mid-inversion. This is the only
  user-visible change; every such combination already crashed, so no working code breaks.
• New public attributes: AbstractMesh.supports_split_regularization,
  AbstractRegularization.is_split_regularization.
• Removed: InterpolatorRectangular._mappings_sizes_weights_split (private).

Downstream impact: none. Every Pixelization(...) call site across all workspaces was
scanned — zero pair a rectangular mesh with a split regularization. Every split usage is on
Delaunay.

Testing

• 24 new tests: all 9 rejected combinations, all 11 still-valid ones, the capability flags,
  the removed pass-through, and the not-a-FitException property.
• PyAutoArray 930 passed · PyAutoGalaxy 1009 passed (unchanged).
• Controls bit-identical: Delaunay + ConstantSplit log_evidence 5096.4420,
  Delaunay + Constant 5084.7513, RectangularUniform + Constant 4779.4288.

Reported by @rhayes777. Closes part of #332.