engine: run top-k CMSHeap on FastPath to match what sketch-bench measures #811

Part of #790.

Motivation

The MILP chooses top-k deployments using sketch-bench measurements of cms-heap-topk-fastpath-vector2d. Its cost and accuracy numbers are only valid if ASAPQuery deploys that same sketch. Today ASAPQuery deploys a different one:

As a result, the planned CPU cost and accuracy describe a sketch the engine does not run.

Agreed scope

Switch ASAPQuery’s top-k accumulator to CMSHeap<Vector2D<i32>, FastPath> directly. This issue is intentionally limited to ASAPQuery:

• Keep both top-k modes. The updater inserts 1 for count; for sum, it rounds the incoming sample value at the engine boundary and passes an i32 increment to the sketch.
• Keep the accumulator integer-only.
• Keep JSON diagnostic output, but remove unused JSON deserialization.
• Make byte serialization explicitly unsupported. Do not preserve or test the portable MessagePack representation.
• Update matrix memory accounting for 4-byte i32 cells.
• Add a focused FastPath regression/parity test and retain existing end-to-end top-k coverage.

Out of scope: Go collector/back-end wire compatibility, planner changes, persisted-state migration, and asap_sketchlib changes.

Done when

• The engine’s CountMinSketchWithHeap runs on FastPath with i32 counters, matching cms-heap-topk-fastpath-vector2d.
• The updater preserves count and sum top-k modes at the engine boundary.
• FastPath behavior and existing top-k end-to-end behavior are covered by tests.