feat: Delaunay-pixelization JIT round-trip support in galaxy→inversion path

[Jammy2211]

Summary

Enables jax.jit(analysis.fit_from) to succeed for models with a Delaunay
pixelization source + Hilbert image-mesh + AdaptSplit regularization. Without
this change the JIT path raises AttributeError: 'NoneType' object has no attribute 'array' inside GalaxiesToInversion.image_plane_mesh_grid_list, because the
Galaxy keys in adapt_images.galaxy_image_plane_mesh_grid_dict are the
original Python objects and the traced Galaxy is a fresh instance produced by
pytree unflatten — identity-keyed dict lookup fails.

API Changes

Behaviour-only change in GalaxiesToInversion.image_plane_mesh_grid_list: when
the galaxy-identity-keyed mesh-grid dict lookup returns None and the dict
contains exactly one entry, that entry is now used by insertion order.

No public signature changes. NumPy path is unaffected (it never hit the fallback).

See full details below.

Test Plan

• pytest test_autogalaxy/ -q --ignore=test_autogalaxy/aggregator — 836 passed
• scripts/jax_likelihood_functions/imaging/delaunay_pytree.py PASSes —
  NumPy and JIT log_likelihoods agree to rtol ~4e-15
• Already-shipped pytree variants (lp, rectangular, mge_group) still pass

Upstream PR

None — this ships as a standalone library change.

Follow-up

The one-entry fallback is a deliberately narrow workaround. It is correct for the
current PoC (single pixelised source) but silently picks the wrong grid when there
are two or more pixelised sources. The principled fix — attach pytree_token to
Galaxy mirroring LightProfileLinear so identity survives pytree round-trips —
is filed at admin_jammy/prompt/autolens/galaxy_pytree_token.md in the workspace
repo and should land before any multi-source-pixelization JIT variant is added to
the queue (e.g. fit_imaging_pytree_delaunay_mge.md — item 10 in the queue).

Full API Changes (for automation & release notes)

Changed Behaviour

• GalaxiesToInversion.image_plane_mesh_grid_list —
  when adapt_images.galaxy_image_plane_mesh_grid_dict[galaxy] returns None
  (because the Galaxy instance has been reconstructed by a pytree unflatten
  cycle and no longer has identity-equality with the dict's keys), the method
  now falls back to the single value in the dict when the dict has length 1.
  Previously this case appended None to the result list, leading to a crash
  downstream when the mesh grid was accessed.

Rationale

• Unblocks the Delaunay-pixelization PoC for jax.jit(analysis.fit_from).
• NumPy path is unchanged — identity lookup succeeds without the fallback there.
• Multi-pixelised-source case intentionally not fixed here; tracked by the
  galaxy_pytree_token.md follow-up prompt.

🤖 Generated with Claude Code

[Jammy2211]

Workspace PR: PyAutoLabs/autolens_workspace_test#38