Accelerate rank-list construction with Numba

[afermg]

Stack

• Base: Add an optional Numba cosine backend #141 (perf/numba-cosine)
• This PR adds only regular-AP rank-list acceleration. Multilabel rank construction remains unchanged.

Summary

When backend="numba" is explicitly selected for regular AP:

• group directed positive and negative scores by query
• sort each class independently
• merge relevance labels using NumPy-compatible key, NaN, and positive-before-negative tie semantics
• avoid the reference path's global concatenation and np.lexsort
• continue using the existing compute.ap_contiguous

The NumPy backend is byte-for-byte unchanged. The optional compiled rank kernel validates a narrow native float32/float64 score contract before JIT dispatch; malformed shapes, mismatched lengths, integer/float16/complex/object/non-native scores, and invalid pair dtypes fail safely and explicitly.

Benchmark

Synthetic JUMP-shaped rank workload: 39,858 profiles, 50,000 positive pairs, 898,101 negative pairs, and 1,896,202 directed entries. Two-run medians:

Backend	Time	Speedup
NumPy reference	18.045 s	—
Numba cold compile	2.938 s	6.1×
Numba warm	0.247 s	73.2×

Rank-list, AP, and null-configuration digests were byte-identical; maximum AP difference was exactly zero.

Separate-process peak RSS was 172 MB for NumPy and 260 MB for Numba because loading/JIT-compiling the optional runtime costs memory despite avoiding global rank-sort temporaries. In normal use #141 has already loaded Numba for cosine, so much of that runtime overhead is shared.

Absolute times are noisy on the contended host, but cold and warm improvements and output parity were repeatable.

Exactness and validation

• 56 focused optional-backend tests passed
• randomized multigraphs and sparse IDs
• duplicates, self-edges, positive/negative overlap
• finite ties, signed zero, NaN, and ±Inf
• positive-only, negative-only, and both-empty inputs
• mixed float32/float64 promotion
• pre-JIT shape, length, pair-dtype, and score-dtype validation
• byte-exact rank labels/counts, AP, configurations, and p-values
• Ruff and diff checks passed
• independent follow-up review: OK