feat(examples): add offline SDK-local per-control timing demo

[josjeon]

Summary

Adds a runnable, offline Python SDK example for the local-control timing work discussed in Part A of the Agent Control load-test thread. It uses the per-control execution_duration_ms added by #283, runs five synthetic built-in control cases (allow, block, steer, SQL, JSON), validates their outcomes, and reports per-control time beside the full SDK-call time. An optional --concurrency mode runs a bounded single-process throughput smoke and records observed calls per second. An HTTP mock rejects accidental network requests and SDK observability is disabled for the run.

This PR is stacked on #283 because that PR supplies the measured field. Retarget it to main after #283 merges.

Verification

• python -m pytest examples/sdk_local_timing/test_demo.py -q: 2 passed using the exact feat(engine): capture per-control execution duration [SAO-18023] #283 SDK environment.
• ruff check examples/sdk_local_timing and ruff format --check examples/sdk_local_timing: passed.
• Ran demo.py --iterations 4 --warmup 2 and --iterations 12 --warmup 2 --concurrency 4: all five cases produced finite, nonnegative control durations and a tabular comparison.
• Repository make prepush could not complete because the configured Splunk Artifactory Python index returned HTTP 401 while resolving google-re2; no test failure was reported before dependency resolution stopped.

Scope

This is a fixture-driven SDK-local timing example. The throughput smoke uses one Python process and event loop, and it does not add an evaluator-only timer, instrument a decorated agent, or establish SDK/server load capacity. Server-side GLoad work is tracked separately in Orbit.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!