planner/engine: deploy top-k CMSHeap on FastPath to match what sketch-bench prices #810

Part of #790.

Problem

The MILP prices top-k with a different sketch from the one ASAPQuery deploys:

sketch-bench's own doc comment on Capability::families notes that fastpath and regularpath have different cost profiles, so the planned CPU cost doesn't match what's deployed. The i32 vs i64 counters also mean the plan undercounts matrix memory by 2×.

Decision

Deploy FastPath in ASAPQuery, so it matches what sketch-bench measures. We are not switching sketch-bench to regularpath.

What the change touches

• The accumulator. asap_sketchlib::sketches::CMSHeap<_, FastPath> is public, so ASAPQuery could build it directly rather than through the portable CountMinSketchWithHeap facade, which hardcodes RegularPath. That needs its own serde, because the facade's MessagePackCodec is what the precompute engine and query engine exchange.
• The Go collector. ASAPCollector/asap-precompute-go emits CountMinSketchWithHeap payloads that ASAPQuery decodes with from_msgpack (see asapedgeprocessor/emit_heap_test.go). A sketch built with one hash path and queried with the other puts keys in the wrong buckets, so the Go producer has to move to FastPath hashing too, or Go-built sketches need to be refused or kept apart.
• asap_sketchlib. The unmerged branches asapv1-only/cmsheap and asapv1-only/cleanup contain 458d02e ("drop the portable MessagePack format"). If that lands, the facade goes away and ASAPQuery has to construct sketches::CMSHeap itself anyway. That is the natural place to make this change, so coordinate with that branch.
• Counter width. Pick one of i32 or i64 and make sketch-bench's wrapper and ASAPQuery agree. With i32, value-weighted top-k (insert_many(round(value))) needs an overflow check.
• Planner. milp_output.rs already maps cms-heap-topk-fastpath-vector2d to CountMinSketchWithHeap. Once the engine runs FastPath, no planner change is needed.

Done when

• The engine's CountMinSketchWithHeap uses FastPath with the agreed counter width.
• Go-produced heap sketches are either hashed the same way or rejected.
• A test checks that a sketch built by the engine and the sketch sketch-bench benchmarks produce the same bucket indices for the same keys.