feat: discover duplicate edges with uncapped IVF radius queries - #10304

Add an alternative to comparing only vectors stored in the same IVF partition: use each live source vector as a query against IVF cells in every active index segment, then emit threshold-matching edges.

Execution flow

Pinned dataset + source (segment, partition)
  → live source row IDs → original source vectors
  → for each source, select IVF cells in every target segment
  → uncapped radius scan over stored vector codes
  → verify candidate distances on original vectors
  → stream canonical (min_row_id, max_row_id, distance)

find_similar_pairs_in_partition discovers edges for one source scope. It searches all target segments, so matches may cross the source partition and segment boundary. Radius scans return every qualifying candidate within the scanned cells; no top-k truncation is applied. Self edges are excluded. Discovery can be asymmetric, so matches are retained regardless of which endpoint found the other; repeated canonical edges may be emitted and can be merged by the caller.

Probe selection and distance semantics

• Automatic probing considers the nearest two centroids by default. Centroid-boundary geometry determines whether an additional cell is within the query's threshold neighborhood.
• Fixed nprobes selects a fixed number of nearest cells. max_nprobes changes the automatic candidate budget; counts must be positive and fixed probes cannot exceed an explicit maximum. Budgets apply independently to each target segment.
• Candidate scans use the index metric/quantizer; final distances are verified on original vectors. This removes candidate false positives, while restricted probes and quantization can still omit matches.
• The scan operates on IVF vector codes independently of HNSW traversal. It does not add HNSW acceleration or claim top-k ANN throughput.

Requires one current-format logical vector index covering the pinned dataset, with stale indexed values rejected. Source/target range blocking is a separate change in #10305.

Coverage includes centroid-boundary matches, more than 100 radius neighbors, nested vector columns and automatic/fixed probe budgets.