feat: nufftax-backed TransformerNUFFT as default; rename pynufft variant to TransformerNUFFTPyNUFFT

[Jammy2211]

Summary

• Replace the pynufft-backed TransformerNUFFT with a JAX-native nufftax-based implementation as the default. Preserve the original pynufft class as TransformerNUFFTPyNUFFT for backwards compatibility.
• The new class matches TransformerDFT to ~1e-13 relative across 13 stress-test configurations (odd/even/non-square sizes, varying pixel scales, sparse/dense uv, low/high frequencies). Pynufft sits at ~6% relative gridding error on production 256×256 cases.
• nufftax supports jax.jit, jax.grad, jax.vmap — unblocks end-to-end JAX-jitted interferometer likelihoods. JIT'd forward NUFFT measures ~20% faster than pynufft (6.1 ms vs 7.7 ms on 256² / 1000 vis).

API Changes

• aa.TransformerNUFFT is now the JAX-native nufftax variant (default).
• aa.TransformerNUFFTPyNUFFT exposes the legacy pynufft variant unchanged.
• Interferometer.apply_sparse_operator() raises NotImplementedError when the dataset's transformer is the new TransformerNUFFT. The sparse-operator path depends on pynufft's kernel-deconvolved adjoint scale; users can opt back via transformer_class=aa.TransformerNUFFTPyNUFFT, or use TransformerDFT (which is what the JAX-likelihood scripts already do).
• pyproject.toml: nufftax added to optional and dev extras alongside pynufft.

Behavior change

TransformerNUFFT.image_from now returns the strict mathematical adjoint of the forward (matches TransformerDFT.image_from exactly). The legacy pynufft class included internal Kaiser-Bessel kernel deconvolution and IFFT normalization — the absolute scale and sign of dirty images differ from before. Use TransformerNUFFTPyNUFFT if you need legacy adjoint behaviour.

Companion PRs

• PyAutoGalaxy: re-export TransformerNUFFTPyNUFFT (forthcoming PR on feature/nufftax-default-transformer).
• PyAutoLens: re-export TransformerNUFFTPyNUFFT (forthcoming PR on feature/nufftax-default-transformer).

These should merge after this PR is merged.

Scripts Changed

None in this repo. Workspace parity script autolens_workspace_test/scripts/interferometer/nufft.py already exists and now reports 0.0 residual between the new class and its inline nufftax helper.

Test plan

• pytest test_autoarray/ — 750 / 750 passed
• Workspace parity script — all four cases pass; new class is bit-for-bit identical to the script's nufftax helper
• Hard-fail guard on apply_sparse_operator: new class raises, DFT and legacy pass through unchanged
• Both classes export through aa.X; Transformer Union extended
• black --check clean

🤖 Generated with Claude Code