asap-planner: --planner milp for SQL #834

Part of #753.

asap-planner --planner milp only accepts --query-language promql (run_milp in asap-planner-rs/src/main.rs). The MILP core (solve_milp, Raqe, workload facts, saturation curves) is language-neutral; PromQL is hard-wired only at the input (ControllerConfig → extract_aqes), the labels (metrics: hints) and the output (promql::generator YAML).

Scope

• --query-language sql only (not elastic_sql).
• Workload facts for SQL are written manually.

Design

• Metric key: table.value_column, so each value column gets its own value_range/shapes in the facts file. Its labels are the table's metadata_columns; MILP hints are built from tables:.
• Requirements extraction: planner-side. Factor the parse/match/top-k detection out of SQLSingleQueryProcessor::get_streaming_aggregation_configs (planner/sql.rs) so the legacy SQL planner and the MILP share it. The engine is untouched: MILP output writes each query's aggregation explicitly.
• Top-k: allowed only when GROUP BY covers every metadata column. The ranked key is then the series identity, which is sketch-bench's model (same as PromQL topk(k, x)); k comes from LIMIT. Otherwise refused: two top-k queries keyed on different columns would both be grouping [] and need different shapes.

  -- supported: ranks (host, region) = whole series
  SELECT host, region, SUM(latency) AS s FROM requests WHERE ... GROUP BY host, region ORDER BY s DESC LIMIT 5;
  -- refused: ranks hosts summed over regions
  SELECT host, SUM(latency) AS s FROM requests WHERE ... GROUP BY host ORDER BY s DESC LIMIT 5;

• AVG: nothing special; it maps to [Sum, Count], which the MILP already rejects (MultipleStatistics).
• Code shape: language-neutral core taking (items, hints), with promql_items(&ControllerConfig) and sql_items(&SQLControllerConfig) front ends. Output keeps the shared roll-up check, dedup and aggregation config; only the final YAML emit and the table_name/value_column fields branch by language.
• Cleanup: NoCleanup, same as MILP PromQL (planner: MILP plan → InferenceConfig should use ReadBased cleanup instead of NoCleanup #800 fixes both). aggregate_cleanup, windowing, sketch_parameters rejected.
• CLI: --clickhouse-url/--clickhouse-database rejected with --planner milp (as --prometheus-url is); metadata_columns must be listed in the config.
• rollup_labels left empty, as in MILP PromQL output.

Verification

• PromQL-twin equivalence: the same workload as SQL on a table and as PromQL on a metric with the same labels and facts. Assert identical deployments, and configs equal apart from the SQL table/value_column fields.
• Engine round-trip: the emitted SQL configs route each query to its planned aggregation (like plan_round_trips_into_engine_configs).

Stages

1. Pure refactor to the neutral core; existing PromQL tests unchanged and passing.
2. SQL front end + SQL output, with the tests above.
3. CLI wiring in main.rs.