perf: devirtualize BVH traversal via templated collide/distance overloads

[facontidavide]

Summary

• Adds templated collisionRecurseT<Node> and distanceRecurseT<Node> next to the existing virtual-dispatch versions in include/coal/internal/traversal_recurse.h.
• Adds templated collide<Node> / distance<Node> overloads in src/collision_node.h that route through the templated recursers.
• Per overload-resolution rules, the templated overload outranks the existing base-pointer non-template (exact-type match beats derived-to-base conversion). Internal callers in collision_func_matrix.cpp and distance_func_matrix.cpp, which already create concrete MeshCollisionTraversalNode<BV> / MeshDistanceTraversalNode<BV> instances, automatically pick up the devirtualised path with zero source changes at the call site.

Implementation notes

• traversal_recurse.h gains collisionRecurseT and distanceRecurseT (header-only templates) that take Node* instead of CollisionTraversalNodeBase*. The 9 virtual calls per BVH node pair (isFirstNodeLeaf, isSecondNodeLeaf, BVDisjoints, leafCollides, firstOverSecond, getFirstLeftChild, getFirstRightChild, getSecondLeftChild, getSecondRightChild, canStop) all become statically-resolved member calls on the concrete Node type — the inliner can fully inline them.
• src/collision_node.h gains collide<Node> / distance<Node> template overloads that wrap the T-prefixed recursers. Front-list propagation and queue-based search paths remain on the virtual codepath (those are rare and not worth the extra instantiations).
• src/collision_node.cpp exports checkResultLowerBound as COAL_DLLAPI so the header-only template can call it without a link-time visibility hole.

Performance impact

Methodology

• Hardware: x86-64 desktop, P-core pinned (taskset -c 4), turbo locked.
• Build: cmake --build build -j12 -DCMAKE_BUILD_TYPE=Release in isolated worktrees; separate libcoal.so per variant. Both base and variant compiled with stock upstream flags (no -march=native).
• Runs: N = 15 interleaved (base, variant, base, variant, …); median reported.
• Workload: coal-test-benchmark (test/benchmark.cpp).

Workload	Before (µs, median)	After (µs, median)	Δ	N	stdev (µs)
coal-test-benchmark total	62591.4	60417.5	-3.47%	15	base 313.8 / variant 324.4

Correctness gate: full ctest suite passes (37/37, including the new coal-devirt_correctness test).

Tests

• The cherry-picked commit already adds test/benchmark_devirt.cpp (a perf benchmark that calls collisionRecurseT directly so the templated dispatch route is exercised at build time).
• Added test/devirt_correctness.cpp — a focused correctness test that, for each MeshCollisionTraversalNode<BV> / MeshDistanceTraversalNode<BV> instantiation (OBB, RSS, kIOS, OBBRSS; OBB has no distance node), runs both the templated path (collide(&node, ...) / distance(&node)) and the virtual-dispatch path (collide(static_cast<CollisionTraversalNodeBase*>(&node), ...)) on identical inputs and asserts the two CollisionResults / DistanceResults are pairwise-equal.
• The whole point of the change is overload-resolution dispatch; a benchmark proves it routes to the templated path, but only this correctness test proves the templated path produces exactly the same result.

Risk & regression analysis

• Overload resolution silently changing: the templated overloads outrank the non-template base-pointer overloads by C++ rules. If a future change introduced an ambiguous overload, the wrong path could be selected silently. The new correctness test catches this — if either path drifts, results disagree.
• Front-list / queue-based variants: not templated. Callers that opt into those still get the virtual-dispatch path. No change in behaviour.
• ABI: no public header signature changes. The added COAL_DLLAPI on checkResultLowerBound is a visibility expansion — safe for consumers.

[lmontaut]

See #858