fix: harden graph and postgres data integrity

[ruvnet]

What changed

• reject unsafe CREATE INDEX CONCURRENTLY HNSW builds before they can duplicate heap tuples
• make ruvector.hnsw_ef_search the canonical GUC, retain ruvector.ef_search as an alias, and apply the configured width on every scan
• bind each PostgreSQL SONA table to its first valid embedding dimension and reject malformed or mismatched trajectories before mutation
• make SONA stats inspect the actual initialized engine rather than silently creating a separate 256-dimensional engine
• repair distributed graph replication serialization for dynamic JSON properties
• bump ruvector-graph to 2.3.1 and ruvector-postgres to 2.0.6

Impact

This prevents duplicate HNSW results and silent SONA training loss, restores operator control over HNSW search cost/recall, and restores the graph distributed/all-features build.

Validation

• cargo test -p ruvector-graph --all-features --lib — 199 passed, 3 ignored
• cargo check -p ruvector-graph --all-features
• PostgreSQL extension cargo check against PostgreSQL 16 with pg16,index-all,sona-learning
• cargo package -p ruvector-graph --allow-dirty --no-verify
• git diff --check

Closes #727.
Closes #728.
Closes #729.
Closes #732.