Benchmarking optimizations

[SK1PPR]

Summary

This PR prepares VortexDB for ANN-Benchmarks integration by improving index configurability, reducing HNSW overhead, and adding Python batch client APIs for faster benchmark data loading.

Changes

• Optimized HNSW distance calculation paths

  • Avoids unnecessary sqrt for Euclidean comparisons
  • Uses normalized cosine vectors with direct 1.0 - dot scoring
  • Simplifies visited-set handling during graph traversal

• Exposed HNSW tuning parameters

  • HNSW_M
  • HNSW_M0
  • HNSW_MAX_LAYER
  • HNSW_EF_CONSTRUCTION
  • HNSW_EF

• Exposed KD-tree rebuild tuning parameters

  • KD_TREE_BALANCE_THRESHOLD
  • KD_TREE_DELETE_REBUILD_RATIO

• Added Python batch client APIs

  • insert_batch(...)
  • search_batch(...)

• Regenerated Python gRPC stubs to include existing batch RPCs

• Kept FlatIndex unchanged since it has no meaningful index-level tunables beyond query-time k and similarity

Testing

cargo test -p index
cargo test -p api
cargo test
cd client/python && UV_CACHE_DIR=/private/tmp/uv-cache-vortexdb uv run --extra dev pytest -q