Fix host dispatch benchmark timing window

[jhinpan]

Summary

• move the final GPU drain outside bench_wallclock's measured interval so it measures host dispatch as documented
• add a deterministic regression test that locks the synchronization/timestamp ordering
• keep the change intentionally narrow; the broader timing-contract and CI analysis is documented in the supporting Gist

Why

The previous implementation sampled the end timestamp after torch.cuda.synchronize(). For GPU-bound workloads that turns the metric into roughly max(host submit, GPU throughput) rather than CPU dispatch overhead, contradicting the function's docstring and making results workload-size dependent.

Test plan

• python3 -m pytest tests/unit/test_launch_overhead.py -q
• python3 -m ruff check tests/unit/test_launch_overhead.py

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[jhinpan]

Superseded by #924, which carries this validated host-dispatch fix and implements the full timing-contract, event-timer cleanup, raw-us schema, verified graph benchmark, and calibrated CI gate.