Implement support for triangulated volumes in simphony

[ggalgoczi]

Summary

Adds OptiX builtin triangle-BVH support to the physics simulation pipeline so volumes flagged via stree__force_triangulate_solid=<lv_name> traverse as triangle GAS instances instead of analytic CSG, mixed inside the same IAS as the analytic CSG GAS.

The TriMesh HitGroupData layout (Binding.h), the triangle build input (sysrap/SOPTIX_BuildInput_Mesh.h), the mesh-side SOPTIX_MeshGroup plumbing and the __closesthit__ch triangle branch were already in place. This PR adds the pipeline + SBT wiring that was missing to actually exercise them.

Changes

feat(csgoptix) commit:

• CSGOptiX/PIP.cc: enable OPTIX_PRIMITIVE_TYPE_FLAGS_TRIANGLE alongside
  _CUSTOM in usesPrimitiveTypeFlags; create a second hitgroup program
  group hitgroup_pg_tri with entryFunctionNameIS = nullptr (OptiX 9 forbids
  an IS program on a triangle hitgroup); add it to optixPipelineCreate and
  optixUtilAccumulateStackSizes.
• CSGOptiX/PIP.h: declare hitgroup_pg_tri alongside the analytic PG.
• CSGOptiX/SBT.cc::createHitgroup: pack each SBT record header with the PG
  matching its GAS kind _ hitgroup_pg_tri when foundry->isSolidTrimesh(gas_idx),
  hitgroup_pg otherwise _ instead of a single PG for all records.

test(ci) commit:

• tests/test_triangulated.sh: runs GPURaytrace on the existing
  opticks_raindrop.gdml twice _ once analytic, once with
  stree__force_triangulate_solid=G4_WATER_solid _ and asserts the
  Opticks: NumHits: counts are bit-identical. A G4Box triangulates to
  12 flat triangles that coincide with the analytic surface, so strict
  equality is the right check.
• .github/workflows/build-pull-request.yaml: invoke the new test next
  to test_GPURaytrace.sh.

Behaviour for runs that don't trigger triangulation

Functionally equivalent to upstream main:

• hitgroup_pg_tri is created but unreferenced (no triangle GAS exists).
• SBT::createHitgroup falls back to hitgroup_pg for every record.
• Slight optixUtilComputeStackSizes overhead from the extra PG, negligible.
• Raindrop regression: 12664 hits in both branches, matches existing tolerance.

Mixed analytic + triangulated

Supported. The CI test itself runs a mixed scene (analytic Pb/AIR/VACUUM + triangulated WATER) in a single IAS. The constraint inherited from the existing CSG bookkeeping is that within a single compound CSGSolid all CSGPrim layers must be either all analytic or all triangulated (see comment at SBT.cc:1021). Across different LVs / GAS indices, mixing works.

Validation

Beyond the CI test, manually validated on a deformed raindrop where G4Box solids were replaced with G4Sphere:

solid	rotation steps	num triangles	hit deficit vs analytic
Box (G4_WATER)	n/a	12	bit-identical, 5/5 seeds
Sphere (G4_WATER)	24 (default)	528	+0.255% (5/5 seeds, ~3_ each)
Sphere (G4_WATER)	48	2208	+0.105%
Sphere (G4_WATER)	200	39600	+0.0086%

Deficit scales as 1/N_ as expected from the inscribed-polyhedron error. N=34 was an outlier (+0.405%) _ G4PolyhedronSphere with prime-factor rotation counts produces non-smooth meshes; multiples of 24 stay on the clean 1/N_ curve.

GPU Simulation time: measured around gx->simulate() + cudaDeviceSynchronize in src/GPURaytrace.h. On a single warm-run measurement at N=200 with 6.08M photons (RTX 4090), triangle BVH was ~10% faster than the analytic CSG-sphere path _ RT cores beat the custom __intersection__is software intersect for the simple sphere; the gap likely flips back to ana for deep CSG trees.

How to enable

export stree__force_triangulate_solid=lv_name_A,lv_name_B
or

export stree__force_triangulate_solid='filepath:$HOME/.opticks/GEOM/${GEOM}_meshname_stree__force_triangulate_solid.txt'

Names are post-StripTail (no 0x_ suffix). Must be set before the G4_Opticks geometry conversion. With no envvar, behaviour matches main.