perf: inspect runtime overhead against sketch and raw exact computation

[zzylol]

Problem and behavior

Before this PR: runtime/thread controls and a reproducible comparison of direct computation, installed-plan execution and HTTP overhead were missing.

After this PR: runtime controls and an inspection executable compare four paths over identical deterministic inputs: direct sketch, raw exact, backend bound-query execution and HTTP. Results must agree before measurements count as successful. Reports retain latency, throughput, request failures/drops, CPU, memory, threads, effective controls and cgroup limits.

Each cell now includes installed-plan.json, six separately measured setup stages, and an explicit list of included/excluded architecture stages. Setup time is separate from request latency. The fixture installs imported state: it does not benchmark Planner search/physical compilation, backend deployment compilation, precompute execution or durable recovery. Listed request stages are not individually timed. See docs/developer_docs/performance/overhead-inspection.md.

The inspection guide also explains how to retain the selected query-time or precomputed Rate/Sum/heap placement, including window width and stride. Runtime controls do not change those Planner decisions.

Stack

Based on #761. #766 and #756 are independent follow-ups; the acceptance stack does not depend on either. This PR does not require diagnostic logging to benchmark the production query path.

Validation and limits

• Strict all-target Clippy passes after merging the current feat(control-plane): scope evidence and compute ERP/analytical workload costs #761 base. The earlier four inspection tests and four-layer smoke run below remain the fixture evidence; the new native Rate/Sum paths are not benchmarked by that synthetic fixture.
• Four-layer smoke run: 10 successful queries per layer, one worker/concurrency, 1,000 samples; each cell retains the installed plan and six setup timings.
• Debug-build checks validate behavior, not a release speedup. The fixture is a DDSketch median over one series and one complete pane.
• Existing Level 3 performance failures are not resolved by this tooling PR.

Refs #758.