Reduce benchmark harness process churn in performance measurements

[spMohanty]

Benchmark measurement currently builds temporary scripts and launches subprocess/perf processes on a per-measurement basis. This adds significant harness overhead for large matrices.

Add batched/persistent execution support so multiple measurement cases can run inside fewer Python/perf processes while preserving existing interfaces and comparability.

Files

• benchmarks/_perf.py

Acceptance Criteria

• Reduce process launch overhead in suite-level runs.
• Keep output semantics for existing benchmark reporting.
• Improve total wall-time for large benchmark sets without changing measured operation definitions.