benchmark: cover multiple testing resampling and unsupervised operations

[TheHiddenObserver]

Goal

Add canonical coverage for maintained multiple-testing/resampling utilities and missing operation-specific unsupervised paths.

Scope

• multiple-testing, p-value combination, bootstrap, and permutation utilities with statistically meaningful metrics;
• standalone NNDescent and operation-specific fit/transform/predict timings for maintained unsupervised estimators;
• aligned cases, seeds, backends, external references, synchronized GPU timing, and explicit failure states;
• structured sources, parser tests, registration, generated assets, and coverage-matrix updates.

Parent: #100