Add CMS-with-heap top-k demo and benchmark vs exact

[zzylol]

Stacked on #20. Adds a Count-Min sketch with heavy-hitter heap (asap_sketchlib::CMSHeap) and a CMS + heap vs exact top-k demo and benchmark, using the same multi-process pinned topology as the KLL demo in #20.

Before → after

• Before: the collector had CMS and CountSketch wrappers but no top-k: CountSketch::top_k returns nothing, and the OTAP carrier accepted only KLL ASAPv1 payloads.
• After:
  • CMSHeapWrapper wraps asap_sketchlib::CMSHeap. It sends ASAPv1 envelopes (kind 0x03 0x00, CountMinSketchWithHeap) and implements FrequencySketch::top_k. Merging re-ranks both heaps on the merged counters. The OTAP carrier now accepts and validates these payloads.
  • The benchmark has two new scenarios, topk-exact and cmsheap, run on keyed traffic: Zipf(1.1) over 2^20 keys.
    • topk-exact: each branch counts keys in a hash map and ships every (key, count) entry; merge sums them.
    • cmsheap: each branch ships one sketch per window (4 x 2048 counters, 10-slot heap); merge merges sketches.
  • The backend checks every window: exact must match, and CMS + heap must reach recall@10 ≥ 0.9.
  • run-demo.sh --scenarios topk-exact cmsheap serves a new Grafana dashboard: both pipelines' current top 10 side by side, recall@10, count error, throughput, and per-process CPU, egress and memory.

Run

demos/kll-grafana/run-demo.sh --scenarios topk-exact cmsheap   # Ctrl+C stops

Open http://localhost:3000/d/asap-topk-stream.

56-core x86 server, 50,000 signals/s per source. Both pipelines keep up. CMS + heap uses 1.7 processor CPU-seconds per million signals; exact uses 13.0. CMS + heap returns the same ten keys in the same order, with counts within 0.6%.

Throughput: CMS + heap vs exact top-k (vs raw forwarding)

65,536-key windows, 3 runs per point, one dedicated CPU per processor, a 56-core x86 server. Median backend-validated signals/s, both sources:

Offered	Raw	Exact top-k	CMS + heap
20,000	21,813	21,813	21,813
50,000	47,988	52,348	47,989
100,000	100,327	100,336	100,335
200,000	196,309	148,316	200,675
400,000	274,848	130,879	397,004
600,000	279,185	130,860	597,659
800,000	283,555	126,516	798,355

Largest sustainable offered rate per source (CAPACITY_DESIGN criteria): 50,000 for exact top-k, 100,000 for raw, and at least 400,000 (the highest tested) for CMS + heap. Above that, exact top-k plateaus near 130,000 signals/s.

Processor CPU and egress summed over branches, merge and estimate, at the same load:

Offered per source	Exact CPU	CMS + heap CPU	Exact egress	CMS + heap egress
10,000	0.29 cores	0.039 cores	43.9 Mbit/s	0.1 Mbit/s
25,000	0.66 cores	0.084 cores	102.9 Mbit/s	0.3 Mbit/s
50,000	1.29 cores	0.165 cores	201.1 Mbit/s	0.5 Mbit/s

Accuracy

The backend checked every window of every run:

• Exact top-k: matched exactly.
• CMS + heap: recall@10 = 1.0, mean relative count error 0.62% over the true top 10. Count-Min only overestimates.

Library pin

This PR moves the asap_sketchlib pin from 010457de to 3fc3442c (current main), because the CMSHeap ASAPv1 payload exists only there (asap_sketchlib#115 and follow-ups). On main, DdSketchState also gained negative and zero stores. The DDSketch wrapper now uses sketchlib's own to_proto/from_proto instead of a local field mapping. All existing sketch, codec and runtime tests pass on the new pin.

Validation

cargo test --features otap-engine,stream-benchmark   # CMSHeap wrapper, carrier round trip, top-k model tests
python3 demos/kll-grafana/verify-streaming.py --scenarios topk-exact cmsheap

All tests pass, along with fmt and Clippy. The demo ran end to end on a 56-core x86 server.