Add continuous KLL window quantiles with Prometheus and Grafana

[zzylol]

Adds a persistent OTAP KLL demo that merges tumbling panes into rolling quantiles and continuously exports p50/p99 to Prometheus, with a provisioned Grafana dashboard.

Before → after

• Before: the finite demo wrote estimates to console and local files, then exited.
• After: the finite demo can export its results; the streaming demo runs until Ctrl+C, exports completed-window estimates repeatedly, and displays tumbling and rolling results in Grafana.
• Windows: retained KLL panes support aligned range queries and configurable rolling lookbacks. Sketches are merged before estimating quantiles; Prometheus stores the resulting gauges.

End-to-end example

From the repository root:

docker compose -f demos/kll-grafana/compose.yaml up -d
cargo build --manifest-path asap-precompute-rs/Cargo.toml \
  --features otap-engine --bin asap-kll-stream-demo

# Run for 30 seconds: 6-second lookback, 1-second panes.
./asap-precompute-rs/target/debug/asap-kll-stream-demo \
  --window-seconds 6 --slide-seconds 1 --run-seconds 30

Open http://localhost:3000/d/asap-kll. Grafana refreshes every five seconds and queries:

request_duration_estimate{quantile=~"0.5|0.99"}
request_duration_rolling_estimate{quantile=~"0.5|0.99"}

Omit --run-seconds for continuous operation. Defaults: 60-second lookback and five-second panes.

Live screenshot using six-second rolling windows and one-second panes.

Validation

python3 demos/kll-grafana/verify-streaming.py \
  asap-precompute-rs/target/debug/asap-kll-stream-demo

Verified repeated writes, stable series, changing rolling values, and matching Prometheus/Grafana queries. Library, window, runtime, and binary tests passed, along with formatting and Clippy. The finite demo export still returns p50=100 and p99=198; Ctrl+C and export failures were checked.

Limits: pane-aligned ranges, approximate KLL quantiles, in-memory retention, and exactly-once complete upstream panes. See the window guide for configuration and semantics.