feat(planner): cost candidates per (sketch_type, params) from sketch-bench's table

[milindsrivastava1997]

Closes #549

greedy_assign now resolves AtomicCosts per candidate instead of one flat cost applied to everything. CMS/HLL/KLL look up sketch-bench's exported table by exact (sketch, params) key; everything else (Sum, MinMax, CountMinSketchWithHeap, Hydra, ...) keeps the old flat stub, with a warn log so that's visible. Table loads at runtime via --atomic-costs (optimizer_cli, candidate-gen-dump) rather than being compiled in.

candidate-gen-dump got the same --atomic-costs flag so it prints resolved costs next to each candidate — using it against sketch-bench's real export (sketch-bench PR #92) is how we found that CMS currently loses to CountMinSketchWithHeap's stub on a Count query, because CMS's real merge_cpu_secs looks ~1000x too high next to its own insert/query numbers (same anomaly flagged in that PR's description, not yet root-caused).

Scope

This is a subset of #524, doesn't close it completely.

• Param grid matched exactly to sketch-bench's grid, lookup by exact key, no interpolation
• Costs are now a function of (sketch_type, params) instead of flat constants — but only for CMS/HLL/KLL
• Table loaded at runtime from a file path, not compiled in
• Loader is a trivial deserialize, no raw-Record math
• "Missing entry -> drop candidate silently" — done, but only within a benchmarked family. Outside that (Sum, MinMax, CountMinSketchWithHeap, Hydra, ...) this does the opposite of Optimizer: param-aware atomic costs from sketch-bench #524's text: keeps the flat stub instead of dropping. That was a deliberate scope call (see issue discussion), not an oversight, but it is a real deviation worth a second look at merge time.
• Not done: CountMinSketchWithHeap's analytic memory bound (heap_size · avg_key_size)

Not done:

• CountMinSketchWithHeap's analytic memory bound (heap_size · avg_key_size).
• The merge_cpu_secs anomaly above, and sketch-bench#14 (HLL memory bug) — both taint the table this depends on.
• subtract_cpu_secs is absent everywhere (asap_sketchlib#69 not done), by design for now.