perf(numba/sizeshape): share the foreground pass across primitives

[timtreis]

Stacked on #78. Internal optimization of the numba sizeshape lane — no behavior change.

Problem

The four primitives in core/numba/_sizeshape.py each recomputed labels_to_offsets(labels) over the full raster independently, so it ran 4× per get_sizeshape call (moments, convex, perimeter, crofton/euler).

Fix

• Add a _Prep NamedTuple wrapping the labels_to_offsets result (lut, n, offsets) — the only prep every primitive shares — plus _foreground_prep(labels).
• The four primitives take an optional prep arg; _sizeshape_2d computes one prep and threads it into all four.
• Each primitive still computes its own prep when called standalone (prep=None), so the public functions and their tests are unchanged.
• The full foreground pixel list (rows/cols/obj nonzero) is needed only by the moment kernel, so it stays inside spatial_moments_2d — standalone perimeter_2d/crofton_euler_2d/convex_area_2d no longer pay for a full-raster nonzero they never read. convex_area_2d does its own nonzero over the boundary mask (sparse — not redundant).

Impact

Removes 3 redundant labels_to_offsets passes: ~9 ms / ~5% on the 1080² / 132-object tile.

Verification

• ruff check + ruff format --check: clean.
• test/test_sizeshape_numba.py: 24 passed — bit-exactness, dispatch, and edge cases unchanged.

🤖 Generated with Claude Code

[timtreis]

Folded into #78. The shared-prep optimization only touches core/numba/_sizeshape.py (a file introduced by #78), so it can't merge to main independently. Commit a4974d0 is now part of #78's branch feat/numba-sizeshape; GitHub auto-closed this PR as merged once that commit landed on its base branch. Branch deleted.