Nanoflann KD-tree spatial indexing for SVA/SRA

[adthoms]

Summary

• Replace O(N^2) brute-force pairwise compatibility checks in MakeVecAvgInvGraph (SVA) and MakeRotAvgInvGraph (SRA) with nanoflann KD-tree radius searches, reducing complexity to O(N log N) build + O(N * k * log N) queries
• For SVA: direct KD-tree radius search in measurement space using squared L2 distance
• For SRA: embed rotations as 9D vectors (flattened 3x3 matrices), use chordal distance as the KD-tree metric; for geodesic mode, use chordal upper bound as a pre-filter then refine with exact geodesic distance check via existing So3GeodesicDist functor
• Add shared BuildGraphFromHalfEdges helper that converts half-edge adjacency lists to any GraphsStorageType (ADJ_LIST, CSR, ATOMIC_CSR)
• Add nanoflann v1.6.2 as a new dependency via download_project (same pattern as xenium)

Test plan

• New SVA KD-tree produces correct results across graph types test validates all three graph storage types (ADJ_LIST, CSR, ATOMIC_CSR) produce identical inlier results
• New SRA KD-tree chordal equivalence test validates chordal distance mode with and without outliers
• New SRA KD-tree geodesic equivalence test validates geodesic distance mode with and without outliers (using CSR storage)
• Existing SVA/SRA tests continue to pass (same API, same expected results)

🤖 Generated with Claude Code