refactor: defer nufftax/jax and numba imports to first use

[Jammy2211]

Summary

Defers PyAutoArray's two heavy import-time costs to first use, as part of the cross-repo import-time task (PyAutoFit#1505): the nufftax import (which pulls jax, ~0.7s) and numba jit decoration (~0.2s). import autoarray drops to ~1.15s with neither jax, nufftax nor numba loading on bare import.

• operators/transformer.py: the module-level import nufftax and the version-gated 0.6.x batcher shim move into an idempotent _load_nufftax(), called from TransformerNUFFT.__init__ and the three methods that dereference the module global (_forward_native, image_from, transform_mapping_matrix) — instances unpickled in multiprocessing workers never re-run __init__, so the method anchors are load-bearing.
• numba_util.py: jit() queues functions and hands them to numba on the first call of any decorated function, rebinding each defining module's global to the real dispatcher so nopython cross-calls (5 exist) still resolve at compile time. ModuleNotFoundError → plain function fallback preserved; config-driven nopython/cache/parallel options unchanged.

API Changes

None — internal changes only. All 29 @numba_util.jit() functions and TransformerNUFFT behave identically; laziness only moves when the underlying imports happen.
See full details below.

Test Plan

• Full suite: 1062 passed
• TransformerNUFFT numpy/jax visibilities agree exactly; transform_mapping_matrix + image_from verified; pickled-instance-with-reset-module-state (worker simulation) works
• numba cross-call path (binned_mask_from → binned_image_from) compiles and runs; module global rebinds to CPUDispatcher
• Downstream autogalaxy (1111) and autolens (538) suites green against this branch

Full API Changes (for automation & release notes)

Changed Behaviour

• autoarray.numba_util.jit — decorated functions are lightweight wrappers until first call (then real numba dispatchers, module globals rebound). Dispatcher-only attributes (e.g. .py_func) are unavailable before first call; no users exist in the organism.
• autoarray.operators.transformer — nufftax (and jax via it) import on first TransformerNUFFT use instead of at module import; a missing nufftax now errors at construction (same nufftax_exception()), not silently at import.

Migration

• None required.

Generated by the PyAutoLabs agent workflow.