Round the grid origin to a whole pixel

[smartalecH]

Alternative to #3298 for #1051, opened as a draft so we can decide which approach is right.

Root cause. grid_volume::set_origin(const vec&) rounds to the nearest half pixel, so any geometry_center that is not a whole number of pixels gives the grid an odd origin io (e.g. 0.1 at resolution 15 is 1.5 px). The Yee lattice assumes io is even — icenter() documents this and center_origin() enforces it — and loop_in_chunks (vec2diel_floor) and grid_volume::interpolate place each component on absolute ivec parities. With an odd io those disagree with where fields and materials are actually stored, so:

• sources are deposited half a pixel from where they are evaluated (materials are not affected),
• get_field_point, DFT monitors and other loop_in_chunks readouts are likewise half a pixel off,
• source deposition depends on the chunk layout, because the chunk-owned ranges no longer tile the source's interpolation lattice.

Fix. Round the origin to a whole pixel, as center_origin() already does, and round Simulation.geometry_center the same way in Python so the two agree. The rounding is documented.

Evidence (2D/3D, all six components; point, line, plane and boundary-straddling sources on/off chunk splits; random non-pixel geometry_center):

	master	master, geometry_center pre-rounded	this PR
split vs. unsplit, 80 random configs	23/80 differ	0/80	0/80
-np 3/-np 4 vs -np 1, 40 configs	—	—	bit-identical
#1051 script vs. translated reference	0.9916 vs 0.753	—	0.741437 vs 0.741436

The middle column is the key point: master's existing source-deposition code is already chunk-independent once io is even, so none of #3298's loop_in_chunks / fix_boundary_sources changes are needed.

Behavior change. A non-pixel geometry_center now snaps to the nearest pixel instead of the nearest half pixel. Such runs previously misplaced sources and monitors relative to the structure.

Tests: the chunk-boundary tests from #3298 (unchanged), plus test_geometry_center_rounded_to_pixel.

🤖 Generated with Claude Code

[smartalecH]

Folded into #3298, which now contains this change (approved in review there).