fix: area magnification source-plane area wrong when over-sampling is capped #686

Found by arc phase 6b (autolens_workspace_test#351). LensCalc._sub_pixel_magnifications_from (PyAutoGalaxy#685) sets the sub-pixel area from the requested sub_size, but Grid2D.from_mask(..., over_sample_size=sub_size) is capped to 2 under PYAUTO_SMALL_DATASETS=1 (over_sample_util.py), so source_plane_area_from comes out (2/sub_size)^2 too small (0.27 of the disk area in smoke at sub_size 4). area_magnification_from is unaffected (the factor cancels).

Fix: derive the sub-pixel area from the sub-pixels actually generated (pixel area × unmasked pixels / sub-pixels). Test with PYAUTO_SMALL_DATASETS=1 set.

Epic cluster-strong-lensing, phase 6a follow-up. Human authorized shipping 6b ("ship it when tests pass", 2026-10-30).