PiPNN kernels: follow-ups deferred from the #1287 review #1439

Follow-ups from the review of #1287 (PiPNN numerical kernels) that did not land in that PR.

• Choose the SIMD group width per architecture. LANES in diskann/src/graph/pipnn/simd.rs is one build-time constant for every architecture. Moving the choice into each architecture needs slice-based grouping (as_chunks::<{A::Vector::LANES}>() needs the unstable generic_const_exprs), a new safety argument for load_group, and an AVX2 comparison of 8 and 16 lanes. Related: Validate PiPNN SIMD width on AArch64/NEON #1423. (review)
• Iterate mask lanes efficiently on aarch64. The top-k kernels visit the set lanes of a comparison through an integer bitmask. aarch64 has no instruction that builds this bitmask directly. A generic mask-lane iterator in diskann-wide could use the best method on each architecture. (review)
• Remove the runtime-k allocation in select_top_k_symmetric. For k > 3, each call collects a Vec<&mut [Candidate]> with one slice per point. A small abstraction over the flat output and k could avoid it. Low priority: the benchmarked configurations use leaf_k 2 or 3. (review)
• Keep large finite L2 distances finite. Leaf L2 writes ‖x‖² + ‖y‖², and GEMM adds -2·xᵀy. Partition L2 does the same with the leader norms. When the norm terms overflow, the result is +inf or NaN, so top-k drops the pair even when the true distance is finite. No test covers large inputs. (review)
• Define empty-input behavior. The kernels document an empty leader set and an empty leaf as unsupported, and partitioning always samples at least one leader. The behavior differs by metric: some metrics panic in row_iter_mut on a matrix without columns. Decide whether these inputs return an error or produce no output.