feat(asap-tools): run the e2e planner through the rqe-optimizer MILP (planner.use_optimizer) #830

experiment_run_e2e and experiment_run_clickhouse start asap-planner without --planner, so they always use the legacy planner. Add a Hydra option to run --planner milp, which uses sketch-bench's rqe-optimizer.

Config (config/config.yaml):

planner:
  use_optimizer: true     # --planner milp
  saturation_dir: null    # remote path; atomic costs read from <dir>/optimizer_cost/rqe_atomic_costs.json (rqe_optimizer::saturation::COST_TABLE)
  workload_facts: null    # remote path to the workload-facts YAML
  w_cpu: null             # null = rqe-optimizer default
  w_mem: null

Behavior when use_optimizer=true:

• ControllerService passes --planner milp --workload-facts --atomic-costs --saturation-dir [--w-cpu] [--w-mem].
• Bare metal uses the remote paths as given.
• Containerized: the compose template adds read-only mounts (/app/saturation, /app/input/workload_facts.yaml). The sketchdb-controller image has to be rebuilt by hand so it includes --planner milp.
• aggregate_cleanup and sketch_parameters are left out of controller_input.yaml, because MILP rejects them (milp.rs:249) and the MILP picks sketch parameters itself.
• These fail fast in Python:
  • saturation_dir or workload_facts is unset
  • controller.punting=true
  • controller.auto_discover_labels=true
  • windowing is set
  • experiment_run_clickhouse is used, since SQL is unsupported by MILP. Pass planner.use_optimizer=false.

Known limitation: MILP output uses NoCleanup (milp_output.rs:159), so optimizer runs keep every aggregate in memory until #800 lands.

Tests:

• tests/test_controller_service.py: CLI args and compose mounts for both modes.
• Validation tests for each fail-fast case, including the ClickHouse script in tests/test_experiment_run_clickhouse.py.

Related: #790, #800.