perf(cosinepair): exact query_row_top_k, halved Theta(n^2) build, zero-copy row views

[Mec-iS]

Fixes #442

Checklist

• My branch is up-to-date with main branch.
• Everything works and tested on latest stable Rust.
• Coverage and Linting have been applied

Current behaviour

• CosinePair::with_top_k builds with an all-pairs ordered double loop: n*(n-1) cosine-distance evaluations regardless of top_k; every evaluation materialises two heap-allocated Vecs through Vec::from_iterator and recomputes both row norms. Measured ~n^3.48 end-to-end on square inputs (CosinePair::with_top_k construction is Theta(n^2) distance evaluations with per-pair allocations; measured ~n^3.48 end-to-end on square inputs #442).
• query_row_top_k always samples only top_k evenly strided candidates (step = n / top_k) with no documentation, and its bounded heap pops the closest candidate when full, so it can return the farthest of the sampled rows.

New expected behaviour

• init() evaluates each unordered pair (i, j) once (symmetric half-scan) and updates the running best of both rows: half the distance evaluations, zero per-pair allocations (zero-copy row views), row norms precomputed once in O(n*d). Distances are bit-identical to Cosine::new().distance(...) (same formula and operation order); measured ~3x faster build on a 1500x64 input (768 ms -> 250 ms, release, Apple silicon). Construction remains Theta(n^2) dot products - top_k does not make it sub-quadratic; docs now state this.
• query_row_top_k is exact whenever approximate is false (the default) and caps the result count at top_k; strided sampling is gated behind CosinePairParameters { approximate: true } and documented as approximate. query() and distances_from() reuse the precomputed norms as well.
• New regression tests: exactness of query_row_top_k vs a brute-force oracle, new() vs with_top_k(n-1) parity (build state and queries), per-row closest neighbour vs brute force, and documented stride sampling under approximate: true. All distance computations still route through the bespoke numbers/ArrayView1 abstractions; no unsafe, no macros.

Change logs

Changed

• Breaking: CosinePair gained a private row_norms field (precomputed row norms). Construct through new / with_top_k / with_parameters instead of struct literals.
• Version bumped to 0.6.11; CHANGELOG.md updated.