fix: make the interferometer simulator's @jax.jit path work

[Jammy2211]

Summary

Follow-up to #420/#421, which fixed the imaging simulator's @jax.jit path and
deliberately excluded the interferometer. Two causes — and one of them was not what
#422 predicted.

The whole change is one file, +56/-1. operators/transformer.py is
deliberately untouched.

1. Interferometer was not a registered pytree

The analogue of #421's site 5: it could not cross the jit return boundary.
Added _register_interferometer_pytrees, mirroring _register_imaging_pytrees —
data and noise_map dynamic; uv_wavelengths, real_space_mask,
transformer, grids, the over-sample sizes and the Nones as aux. The split was
taken from vars(dataset), not the __init__ signature, which is what got the
Imaging split right in #421.

It must also register Visibilities and VisibilitiesNoiseMap. On the imaging
path Array2D is already registered elsewhere; nothing registers these, so they
surfaced as bare leaves:
returned a value of type Visibilities ... at output component [0].

2. A caller that never threaded xp — one line

via_image_from called transformer.visibilities_from(image=image) with no
xp=xp.

#422 got this wrong, and it is worth correcting explicitly. That issue said
TransformerNUFFT._forward_native "hard-converts to NumPy" and needs
"restructuring with its own correctness surface". It does not.
_forward_native already has a complete, jittable JAX branch (lax.scan +
dynamic_slice). The failure reported at transformer.py:660 was in the
NumPy branch — the traceback pointed at the symptom while the cause sat one
frame up, in a caller that never passed xp. Exactly the root shape as site 1 of
#421.

Had I edited the transformer as the issue proposed, I would have restructured
working code to fix a bug that was somewhere else.

TransformerDFT had been passing by luck: its arithmetic flows through tracers, so
the missing xp never showed. Only NUFFT, which calls into _nufftax and then
converts, exposed it.

Verified

• DFT and NUFFT, numpy-eager vs jax-jit, agree to ~1e-11 on real and imaginary parts (complex visibilities, so both are checked)
• Both NUFFT chunk branches under jit match NumPy to ~5e-13 — single-shot and the lax.scan path. chunk_size is a TransformerNUFFT.__init__ arg the simulator never sets, so that branch is unreachable via the simulator and was exercised on the transformer directly rather than skipped
• the returned Interferometer keeps uv_wavelengths, real_space_mask and its noise map — not a broken pytree husk
• NumPy path still ndarray-backed
• test_autoarray 929 · test_autogalaxy 1009 · test_autolens 488 — all identical to pre-change counts

Unit tests are NumPy-only by policy across the stack, so the JAX behaviour is
covered by the parity checks above rather than by test_autoarray/.

Out of scope

TransformerNUFFTPyNUFFT — the legacy pynufft backend is not JAX-traceable and is
not expected to be.

Docs follow-up (after this merges)

The interferometer __JAX Variant__ sections in both workspaces still say the
jitted wrap "does not currently work" (autolens_workspace#379), and the
TransformerNUFFT "supports jax.jit" claim in
autolens_workspace/scripts/interferometer/simulator.py is currently unqualified.
Once this lands, do for interferometer what autolens_workspace#381 did for
imaging: restore the recipe, uncomment the call, and add the script to
smoke_tests.txt so CI executes it.

Generated by the PyAutoLabs agent workflow.