Add the models namespace, port basic-voxel (209 → 233 bindings)

[burinc]

Phase 3's biggest item. Adds raylib.models — 24 new bindings (209 → 233)
plus the five structs they need — and ports basic-voxel as the end-to-end
proof. 106 → 107.

This is the largest single addition to the binding surface so far, and it
opens the models examples, which are the biggest unported group (21 of them).

The part that matters

The Model struct changed between raylib 5.5 and 6.0. 5.5 ended with
boneCount/bones*/bindPose*; 6.0 replaces those with a nested
ModelSkeleton, a currentPose pointer and a boneMatrices pointer.

Model is passed by value to every draw call, so a wrong layout doesn't
error — it reads garbage. Binding this before the 6.0 upgrade would have
produced code that silently broke on it. Worth noting the 6.0 upgrade turned
out to be a prerequisite for this work rather than a parallel nicety.

Layouts verified two ways. By size against hand-computed offsets (mesh
120, material 40, material-map 28, model-skeleton 24, model 136), then against
real raylib data — which is the check that actually proves the offsets:

• a generated 2×2×2 cube reports 24 vertices, 12 triangles, nonzero VAO
• the model it loads into reports meshCount 1, materialCount 1, and validates
• its bounding box is exactly −1..+1
• its transform reads as an exact identity matrix — 16 floats can only all
  land correctly if every preceding offset is right

Pointer fields are real ::mem/pointers with explicit padding rather than the
split lo/hi ints used for Shader. coffi doesn't insert alignment padding and
fails loudly at alias definition if a pointer is misaligned, so every :_pad
is load-bearing.

set-model-material-color!

raylib has no function for this — its own examples reach into
model.materials[i].maps[j].color directly, so this walks the same pointers.
It multiplies with the tint passed to draw-model! rather than replacing
it, which is why basic-voxel sets both to BEIGE: setting only the tint
renders the cubes noticeably lighter than the C does.

Struct moves

Matrix is new in raylib.structs. BoundingBox and Camera3D move there
from core/collision and core/camera3d, so raylib.models doesn't depend
on those namespaces for types it needs by value. core/camera3d keeps a local
alias so existing ::camera3d references still resolve, and all nine
existing models examples were re-run to confirm no regression.

basic-voxel

pick-voxel takes its ray/box test as an argument rather than calling raylib,
so the nearest-hit rule is checkable without a window: nearest of four wins,
no hits gives nil rather than removing something arbitrary, an empty world
gives nil rather than erroring, a tie returns a set member.

Gate: 137 namespaces, 0 lint errors, 69 warnings.

What this unlocks next

Three examples are now fully clear, and several need only a small tier more:

next step	unlocks
LoadModelAnimations / UpdateModelAnimation / UnloadModelAnimations	5 examples
LoadImage / LoadTextureFromImage / UnloadImage + GenMeshCubicmap/GenMeshHeightmap	2 examples
GetRayCollisionMesh	mesh_picking

One decision for you. billboard_rendering is fully unblocked and its
asset (billboard.png) is CC0. directional_billboard is also unblocked but
needs skillbot.png, which — like patterns.png — has no licence entry
upstream. I've held both rather than add a second unattributed asset while
the first is still an open question.