Add a bias-controlled large-N benchmark runner

[michael-denyer]

What changed

• adds an interleaved A/B stage benchmark for kinship, eigendecomposition, rotation, and Mode 4
• verifies bitwise-identical numerical outputs before reporting timings
• rebuilds workers for each ABBA/BAAB block and reverses creation order to remove allocation and process-layout bias
• documents use and adds regression coverage for the ordering protocol

Evidence

• focused regression tests: 2 passed
• full suite: 2,295 passed, 6 skipped, 3 xfailed
• full repository quality gate passed

Measurement note

A first 3,000-sample eigensolver comparison showed a stable 3-4% apparent gap between source-identical workers. Alternating worker creation order removed that false signal. The corrected result crosses zero and is reported as inconclusive. The runner is intended for comparisons on the Databricks ILP64/MKL large-scale machines, where the expected change appears near 100k samples.