feat: add Bvh::cast_ray_best for SIMD-accelerated generic ray queries

[zmerlynn]

Problem

CompositeShapeRef::cast_local_ray_and_get_normal calls bvh().find_best() directly with BvhNode::cast_ray as the AABB cost function. This always uses the non-SIMD ray-AABB test, even on f32/dim3/simd-is-enabled builds where cast_inv_ray_simd is available.

Meanwhile, the sibling method cast_local_ray calls bvh().cast_ray() which does dispatch to the SIMD path. So TriMesh::cast_local_ray gets SIMD but TriMesh::cast_local_ray_and_get_normal does not — a surprising inconsistency.

The root cause: Bvh::cast_ray returns Option<(u32, Real)>, so it can't be used when the caller needs RayIntersection (which includes the surface normal). The caller was forced to use find_best directly and provide its own (non-SIMD) AABB cost closure.

Solution

Add Bvh::cast_ray_best<L: BvhLeafCost> — identical to cast_ray but generic over the return type. Two #[cfg]-gated variants provide the same SIMD/non-SIMD dispatch as cast_ray:

// Non-SIMD
pub fn cast_ray_best<L: BvhLeafCost>(
    &self, ray: &Ray, max_time_of_impact: Real,
    primitive_check: impl Fn(u32, Real) -> Option<L>,
) -> Option<(u32, L)>

// SIMD (f32 + dim3 + simd-is-enabled)
pub fn cast_ray_best<L: BvhLeafCost>(
    &self, ray: &Ray, max_time_of_impact: Real,
    primitive_check: impl Fn(u32, Real) -> Option<L>,
) -> Option<(u32, L)>

RayIntersection already implements BvhLeafCost (returns time_of_impact as cost), so cast_local_ray_and_get_normal can now simply call bvh().cast_ray_best(...).

Test plan

• cargo test — all 520 tests pass across parry2d, parry2d-f64, parry3d, parry3d-f64
• cargo fmt -- --check
• RUSTFLAGS="-D warnings" cargo clippy
• RUSTDOCFLAGS="-D warnings" cargo doc --workspace --no-deps
• Doc-test example on cast_ray_best with proper feature gate
• Benchmark comparison (SIMD vs non-SIMD) — would appreciate guidance on parry's benchmarking setup if one exists

🤖 Generated with Claude Code