Bound parallel worker counts

[afermg]

Summary

Introduce one explicit, memory-aware worker budget across similarity, AP, null-distribution, and mAP aggregation paths:

• default cap of 8 workers
• affinity-aware available-CPU detection
• task-count cap
• max_workers= keyword controls on public AP APIs
• COPAIRS_MAX_WORKERS process-level override
• serial execution that still honors progress reporting
• consistent validation for empty and nonempty task sets

Benchmark

Synthetic cosine workload: 5,000 × 196, 10,000 pairs, 20 batches, identical output digest:

	origin/main default	This PR default
Resolved workers	20	8
Wall time	0.0323 s	0.0226 s
Peak RSS	162 MB	132 MB

This is 1.43× faster with 18.5% lower peak RSS for the benchmark.

The default of 8 is based on a real JUMP-lite scaling sweep: 1/4/8/16 workers used approximately 459/507/862/1,415 MB; 8 workers retained most throughput while avoiding the multi-GB pressure of higher defaults. Explicit and environment overrides remain available.

Validation

• 48 focused tests passed
• worker resolution, affinity, environment parsing, empty tasks, progress at one worker, AP/multilabel/null/mAP propagation, and exact output parity covered
• Ruff and git diff --check passed
• independent follow-up review: OK