PiPNN 7/7: add vertical partition Top-K

[weiyaoluo (SeliMeli)]

• Does this PR have a descriptive title that could go in our release notes?
• Does this PR add any new dependencies? No.
• Does this PR modify any existing APIs? No public API changes.
• Is the change to the API backwards compatible? Yes.
• Should this result in any changes to our documentation, either updating existing docs or adding new ones? No.

Reference Issues/PRs

Stack parent: #1295

What does this implement/fix? Briefly explain your changes.

Purpose

Replace Partition's leader-lane ranking with a vertical Top-K kernel. Each SIMD lane owns one point and maintains an independent runtime-width Top-K while leaders are scanned in sampled order.

This PR does not include the fixed-K POC paths, feature toggles, benchmark code, Leaf changes, stripe tuning, or f32 leader IDs.

Data flow

1. Partition GEMM writes leader-major dots as [leader, point].
2. A complete SIMD group loads one leader's scores across point lanes.
3. Runtime rank vectors keep score and exact u32 ID state lane-for-lane.
4. Reusable worker scratch owns the O(K × SIMD-width) vectors and row-major ID scatter buffer.
5. A scalar tail handles points after the complete SIMD prefix.

The metric seam supplies matching vertical formulas for L2, Cosine, normalized Cosine, and inner product.

Invariants

• Runtime K has no fixed ceiling.
• Strict comparisons preserve sampled-leader order for equal scores.
• NaN and positive infinity do not displace rankable candidates.
• Finite f32::MAX remains rankable.
• Unfilled ranks retain u32::MAX and never become indexed leader IDs.
• Score and exact-ID vectors use the same compile-time lane policy.
• No per-point allocation occurs; worker scratch grows and is reused.

Review order

1. diskann/src/graph/pipnn/simd.rs: score-lane to exact-ID-vector mapping and mask conversion.
2. diskann/src/graph/pipnn/kernel_metric/partition.rs: leader-major metric formulas.
3. diskann/src/graph/pipnn/partition_kernel.rs: vertical insertion, ID scatter, and scalar tail.
4. diskann/src/graph/pipnn/partitioning.rs: architecture-typed worker scratch ownership.

Validation

• cargo test -p diskann --features pipnn graph::pipnn::partition_kernel — 40 passed.
• cargo test -p diskann --features pipnn — 649 passed; doctests passed.
• Temporary one-line f32x16 to f32x8 policy switch — 40 focused tests passed; source restored.
• Native V4 focused run — 40 passed.
• cargo clippy -p diskann --features pipnn --all-targets -- -D warnings — passed.
• AArch64 and Windows GNU cross-target cargo check — passed.
• cargo fmt --all --check, git diff --check, LSP diagnostics, and lens diagnostics — passed.

No performance claim is made in this PR. Representative end-to-end profiling remains a follow-up gate before claiming a build-time improvement.