Compute paired SciPy distances row-wise

[afermg]

Summary

Replace the full B × B scipy.spatial.distance.cdist allocation used only to read its diagonal with direct row-wise kernels for:

• Hamming
• Jaccard
• squared Euclidean
• Euclidean
• city-block / Manhattan

Unsupported metrics and callables retain the SciPy fallback. Public batched similarity results remain float32; the private helper retains SciPy's float64 boundary.

Benchmark

Synthetic paired rows, 1,000 × 196, two repeats, median wall time:

Metric	origin/main	This PR	Speedup
Hamming	3.653 s	0.000198 s	18,430×
Jaccard	6.156 s	0.000538 s	11,440×
squared Euclidean	2.432 s	0.000849 s	2,865×
city-block	5.310 s	0.000884 s	6,004×

Hamming and Jaccard outputs were byte-identical. Squared-Euclidean and city-block results matched within floating-point tolerance. At 4,000 × 196, Hamming peak RSS dropped from 235.5 MB to 102.2 MB while producing an identical digest.

The host was heavily contended, so absolute timings are noisy; the removal of quadratic allocation/work is deterministic.

Validation

• 47 focused tests passed under SciPy 1.16.3 and cached SciPy 1.13.1
• signed/unsigned, boolean, float, nonfinite, empty-shape, alias, callable, and public-dispatch coverage
• Ruff and git diff --check passed
• independent review: OK with non-blocking notes