Optimizer: cost candidates per (sketch_type, params) via sketch-bench's atomic-cost table #549

Carved out of #524

greedy_assign resolves AtomicCosts per candidate instead of applying one flat cost to every candidate in a run:

• CMS/HLL/KLL candidates look up sketch-bench's exported atomic-cost table by exact (sketch, params) key — matches candidate_gen.rs's param grid exactly, no interpolation.
• Every other AggregationType (Sum, MinMax, CountMinSketchWithHeap, Hydra, ...) keeps the existing flat stub (AtomicCosts::default()), logged with a warning so the fallback is visible rather than silent.
• The table loads at runtime from a file path (--atomic-costs on optimizer_cli and candidate-gen-dump), not compiled in. The loader is a trivial deserialize — no raw-record math on the ASAPQuery side.
• candidate-gen-dump --atomic-costs <path> prints each candidate's resolved cost, labeled [real]/[stub], for auditing.

Deliberately not in scope here (tracked on #524 instead):

• CountMinSketchWithHeap's analytic memory bound (heap_size · avg_key_size) — still rides the flat stub.
• Data quality of the table itself: sketch-bench's real merge_cpu_secs for CMS looks ~1000x too high next to its own insert/query numbers (found via candidate-gen-dump, currently makes CMS lose to CountMinSketchWithHeap's stub on Count queries) — see sketch-bench PR VictoriaMetrics docker service orchestrates a compose file with multiple containers, and thus breaks the semantics of functions like get_container_name or get_health_endpoint #92's description. Also sketch-bench#14 (HLL memory bug) and asap_sketchlib#69 (subtract unimplemented, so subtract_cpu_secs is absent from the table entirely).