Make held-out evaluation paired or alternating across strategies

[lullu57]

Summary

Held-out evaluation is performed sequentially by strategy immediately after each strategy finishes calibration, rather than in a paired or alternating order across the competing final configurations.

Evidence

• src/kernel_tuner/experiments/orchestrator.py:576
• src/kernel_tuner/experiments/orchestrator.py:589
• src/kernel_tuner/experiments/orchestrator.py:618
• docs/04_experiment_protocol.md:116

Why This Matters

Sequential held-out evaluation is exposed to time-order effects such as thermal drift, cache state, and other host/GPU noise. That weakens the fairness of final comparisons.

Expected Fix

• Rework held-out evaluation to benchmark competing selected configs in a paired or alternating order on the same held-out shapes.
• Record the held-out measurement order explicitly.
• Add tests for ordering behavior where feasible.

Acceptance Criteria

• Held-out evaluation uses a shared ordering policy across compared strategies.
• Measurement ordering is persisted for auditability.
• The protocol and implementation match.