perf(decorators): to_grid propagates only an explicit over_sampled; Grid2D.over_sampled short-circuits at sub-size 1 (#514)

[Jammy2211]

Summary

Phase 1 of the numpy-deflections-cpu epic (#514). Two zero-numerics fixes on the grid-decorator path that every mass-profile deflection call goes through:

1. GridMaker.via_grid_2d no longer materialises the over-sampled grid of an already-wrapped Grid2D. It propagated getattr(result, "over_sampled") / getattr(result, "over_sampler"); when the decorated body returns a Grid2D (the chained @to_grid methods of every spherical mass profile in PyAutoGalaxy) those getattrs fire the lazy properties, which run a per-pixel Python loop (over_sample_util.grid_2d_slim_over_sampled_via_mask_from: two np.linspace + one np.meshgrid per pixel) on every call. The value built that way was also mask-derived, so it ignored the profile's centre and rotation; nothing ever read it. Both sites now read the private _over_sampled / _over_sampler, so only a value the caller explicitly set propagates (the two load-bearing callers, Galaxy.traced_grid_2d_from and Tracer.traced_grid_2d_list_from, set it explicitly and round-trip unchanged).
2. Grid2D.over_sampled short-circuits at uniform sub-size 1 to a copy of the slim grid as a Grid2DIrregular (the loop's output at sub-size 1 is the slim grid in the same order; with a non-zero mask origin the two differ by at most 1 ULP, 2.8e-16). This skips the same loop for every sub-size-1 grid, e.g. the pixelization grid of every CPU likelihood evaluation.

Measured with the new autolens_profiling/scripts/lens/deflections/ cells (hst, 15,361 pixels, OMP_NUM_THREADS=1, direct Grid2D call, before → after): IsothermalSph 699 ms → 0.92 ms, PowerLawSph 701 ms → 1.37 ms, NFWSph 591 ms → 5.8 ms, gNFWSph 1.20 s → 349 ms (the remainder is the MGE, phase 2). Elliptical profiles were never affected (their bodies return a bare array). Every deflection pin passes at rtol 1e-6. Note the direct-Grid2D penalty did not reach the likelihood's ray-trace, which wraps each plane in Grid2DIrregular first; the likelihood-side gain of this phase is the companion tracer change (PyAutoLens / PyAutoGalaxy PRs).

API Changes

None — internal changes only. Grid2D.over_sampled at uniform sub-size 1 now returns a copy of the slim grid (a Grid2DIrregular, as before) instead of the loop's identical output.
See full details below.

Test Plan

• test_autoarray: 1362 passed (4 new). Negative pin test__to_grid__does_not_materialise_over_sampled_of_wrapped_grid control-tested: fails on main, passes here.
• Positive pin: an explicit over_sampled= sentinel survives via_grid_2d.
• Grid2D.over_sampled at sub-size 1 equals the slim grid; at sub-size 2 it has 4× the points and differs.
• autolens_profiling/scripts/lens/deflections/{total,dark}.py --instrument hst: pins PASSED (rtol 1e-6), timings above.
• CI green; downstream PyAutoGalaxy / PyAutoLens suites (run in the task worktree against this branch: 1158 / 576 passed).

Full API Changes (for automation & release notes)

Changed Behaviour

• autoarray.structures.decorators.to_grid.GridMaker.via_grid_2d — propagates only an explicitly-set _over_sampled / _over_sampler from a wrapped Grid2D result; never triggers the lazy properties.
• autoarray.structures.grids.uniform_2d.Grid2D.over_sampled — at uniform over_sample_size == 1 returns a Grid2DIrregular copy of the slim grid without running the per-pixel over-sampling loop (≤ 1 ULP from the loop's output for a non-zero mask origin; identical otherwise).

Generated by the PyAutoLabs agent workflow. Epic ledger: PyAutoMind/draft/feature/autogalaxy/numpy_deflections_cpu_speedup.md. Companion PRs: PyAutoGalaxy + PyAutoLens (tracer double trace at sub-size 1), autolens_profiling (scripts/lens/ measurement package + before/after).