Precompute cosine normalization for indexed pairs

[afermg]

Summary

For the exact distance="cosine" AP paths:

• gather and normalize referenced feature rows once
• reuse the prepared rows for positive and negative indexed pair dot products
• avoid repeatedly normalizing copied pair batches
• preserve exact multiply-then-sum arithmetic, float32 output, tie ordering, and generic/custom-distance fallback behavior

Only referenced rows are normalized, preserving prior behavior for unpaired zero/nonfinite rows. C-, Fortran-, and strided inputs are gathered into the same C-order arithmetic layout as the previous batched implementation.

Benchmark

Synthetic 5,000 × 196, 100,000 pairs, four workers:

	origin/main	This PR
Wall time	98.74 s	74.84 s including one-time preparation
Peak RSS	596 MB	418 MB
Output	SHA-256 identical	SHA-256 identical

That contended run measured 1.32× and 30% lower RSS. An earlier real JUMP-lite worker sweep on the same implementation measured 1.3–4.7× kernel speedups depending on batch/worker count, with exact checksums. A current real end-to-end run retained the exact AP digest and reduced peak RSS from 2.86 GB to 1.61 GB, although host contention obscured wall-time improvement.

Validation

• 47 focused compute/AP tests passed
• exact float32/float64 parity for C/Fortran/strided inputs and ties
• referenced/unreferenced zero and infinity behavior covered
• regular and multilabel paths covered
• empty pairs and generic metric/callable fallback covered
• Ruff and git diff --check passed
• independent follow-up review: OK