Fix matched-budget accounting for benchmark and profile measurements

[lullu57]

Summary

SelectionBudget.max_benchmarks and SelectionBudget.max_profiles are currently enforced per config_id, but each broker request expands across multiple calibration/profile shapes. This makes the recorded budget usage smaller than the real number of measurements and breaks the matched-budget contract.

Evidence

• src/kernel_tuner/selector/engine.py:560
• src/kernel_tuner/baselines/strategies.py:143
• src/kernel_tuner/experiments/orchestrator.py:327
• docs/04_experiment_protocol.md:42
• docs/04_experiment_protocol.md:138

Why This Matters

The paper framing depends on fair matched-budget comparisons. Right now a budget of K can become K * num_shapes actual runtime or profile measurements while the system still reports only K.

Expected Fix

• Redefine broker request interfaces or accounting so benchmark/profile budgets are consumed per measurement, not per config.
• Ensure selector and baselines consume budgets under identical semantics.
• Update reported benchmarks_requested and profiles_requested to match real measurement counts.
• Add tests that fail if one config request silently expands beyond the budget.

Acceptance Criteria

• A runtime benchmark request for one config over N calibration shapes consumes N benchmark budget units.
• A profile request for one config over N selected profile shapes consumes N profile budget units.
• Decision artifacts report true budget usage.
• Tests cover selector and baseline paths.