Expose the coplanar-group map of the input triangles

[danielepanozzo]

What

out_triangle_provenance is indexed by coplanar group, and nothing in the API said which group an input triangle landed in. A caller therefore received per-group lists of output faces with no way back to the input — and so no way to carry a per-triangle attribute (a material tag, which input file a surface came from) through the arrangement.

The mapping was already computed: BSPcomplex::constraint_coplanar_group, filled by computeCoplanarGroups(), and even written out by the -g debug dump. It was simply never returned.

This adds one output to embed_tri_in_poly_mesh:

std::vector<uint32_t>& out_triangle_group,   // per input triangle -> index into out_triangle_provenance

Why it is composed, not returned raw

Constraint indices are not input triangle indices. read_nodes_and_constraints drops degenerate (misAlignment) triangles as it copies them, compacting everything after them, and records the original index in tri_original_index. So the new output composes the two maps and hands back the caller's own numbering:

• a triangle dropped as degenerate maps to UINT32_MAX — it contributed nothing to the arrangement and belongs to no group;
• edge- and point-forcing triangles are grouped internally but carry no original index, so they never name a slot.

Tests

tests/unit_tests_3d.cpp is new — the first unit tests of the 3D embedding API that CI actually runs. embed_regression covers a real model but is hidden behind Catch2's [.] tag because it takes over an hour; these use a six-tet cube and four triangles and finish in well under a second.

They pin the three cases that matter:

input	expected
two triangles sharing an edge, exactly coplanar	same group
a third sharing an edge but folded out of plane	different group
a degenerate (collinear) triangle	UINT32_MAX, later triangles still on their own input indices

The degenerate triangle sits in the middle of the input on purpose: with it last, a map that returned the internal constraint numbering would pass by coincidence. Verified as a negative control — reverting the composition to the identity makes the test fail.

embed_regression also gained a cheap invariant check on the new output.

Local ctest: full suite green.