fix(graph): reject non-finite typed vectors and ranking scores

[rudycelekli]

Problem

Typed graph search accepted NaN/Infinity query components, typed node writes accepted non-finite embeddings, and a legacy invalid embedding could enter top-k ranking. The cited typed_graph.rs partial_cmp().unwrap() in #931 is in a test reference sorter, but these are user-reachable non-finite vector failures in the same path.

Change

• Validate finite query components and typed node embeddings before search/write, including hybrid text queries and overflowing cosine query norms.
• Ignore legacy/directly inserted invalid embeddings when scoring or building an ANN index; also discard non-finite scores from arithmetic overflow.
• Use total_cmp for typed graph result and reference ordering.
• Add regressions for NaN/Infinity inputs, legacy data, hybrid embedder output, and score overflow.

Related to #931. This fixes the typed graph slice, not the wider workspace sweep.

Verification

• Before fix, search_rejects_non_finite_query_vectors failed because a NaN query was accepted; non_finite_stored_vectors_cannot_displace_finite_hits failed because a NaN node was accepted.
• cargo test -p ruvector-graph passed (unit, integration, and doc tests; existing ignored tests remain ignored).
• cargo build -p ruvector-graph passed.
• cargo clippy -p ruvector-graph --lib -- -D warnings passed.
• cargo fmt --all -- --check and git diff --check passed.