fix(asap-tools): complete ClickHouse runs and clean CPU for random E2E pilots

[milindsrivastava1997]

Follow-up to #879, from running workload 01's five-system pilot.

Problems the pilot hit:

• In a fed ClickHouse run, remote_monitor matched processes by clickhouse, which also matched the paced feeder's --sink clickhouse on the same node. The feeder's CPU was added to ClickHouse's, and when the feed ended, the monitor took it for ClickHouse exiting and stopped the query client mid-run. Raw ClickHouse lost about 60% of its queries.
• The run's end was tied to the feed's duration, not to its queries.

Fixes:

• remote_monitor matches the ClickHouse server only, with the pattern clickhouse.server (Docker and bare metal).
• On SIGTERM, paced_feeder stops at the next tick and still writes --stats-out. Achieved rows/s now spans the ticks actually sent.
• A fed ClickHouse run stops its feeders once its queries are done. The feed's duration only bounds a hung run, and the random-workload configs set it to 10× the query phase for every arm. The feeder runs on the next node, as in experiment_run_e2e.py, so it no longer takes ClickHouse's cores. Fed runs now need num_nodes >= 1.
• compare_arms reads a stopped client's truncated results and splits each arm's CPU-seconds into warm-up and query phase.
• The engine test e2e_avg_of_per_series_quantile_over_time gets its own port. It clashed with a test added on main.

Testing:

• experiments: python3 -m unittest discover -s tests -p "test_*.py", 199 pass
• paced-feeder: cargo test, 39 pass
• query engine: cargo test --test e2e_precompute_equivalence, 22 pass (3 runs)
• SIGTERM, checked live against a stub sink: the feeder stops after the tick in flight, writes stats for the ticks sent, and exits 0.

🤖 Generated with Claude Code