[amarcozzi]

Closes #49.

What

Makes attenuate_all cast shadows at the physical solar angle for the voxel
dimensions in env.voxel_dim, instead of assuming cubic voxels.

Why

attenuate_all rotates grid-index coordinates so the sun points straight
down, then floors onto a 1×1 lattice for the shadow sweep. Index space is only
physically proportioned when the voxel is cubic, so voxel_dim never entered the
geometry — it only fed the scalar path length d
(_calculate_d). For any non-cubic voxel
(e.g. a canopy 2 m wide but 1 m tall) the horizontal shadow displacement was
tan(zenith) cells per vertical cell regardless of the true dz/dx ratio, so
shadows landed in the wrong cells.

How

Scale the index coordinates by the voxel aspect ratio, normalized by the
x-resolution (voxel_dim / dx), before rotation. The rotation then acts on
physically-proportioned space, so the shadow angle is correct. Applied to the
leaf, terrain, and sensor rotation inputs; outputs stay in index coordinates.

• Cubic voxels and voxel_dim=None normalize to (1, 1, 1) → no scaling, no
  change.
• An anisotropic voxel is stretched to physical proportions → correct angle.
• When dx != dy the shadow angle is still physical; only the sweep's cell
  buckets become non-square (documented in the docstring).

Tests (TDD)

Written before the implementation — test_voxel_dim_sets_shadow_geometry: a
single elevated occluder over two ground sensors, 45° sun in the x–z plane. The
shadowed ground cell must sit at offset z_top · (dz/dx) cells:

• None, (1,1,1), (2,2,2) → offset z_top (unchanged),
• (1,1,3), (2,2,6) → offset 3·z_top.

Before the fix the cubic/None cases passed and the anisotropic cases failed
(5 != 9); after, all pass. Full suite: 56 passed — the stored-result
assert_equal tests confirm the cubic/None path is byte-for-byte unchanged.

⚠️ Breaking change

Results change for any non-cubic voxel_dim (they were previously wrong).
Cubic-voxel and voxel_dim=None results are unaffected.

Scope / interaction

Kept to attenuate_all (the path the FastFuels treevox consumer uses);
attenuate_surface is unchanged. Independent of #48/#50 (origin offset) but
they touch the same three rotation lines, so expect a small merge conflict — the
combined form is (coords + origin_offset) * voxel_scale, which is compatible
(the global offset is applied before scaling, preserving tile invariance).

[amarcozzi]

Some qualitative evidence of this occuring