Detect distribution drift against each series' recent windows

[zzylol]

Part 5 of the RFC 0001 stack: use case 2, distribution drift and baseline comparison.

The detector keeps each quantile series' last 5 KLL windows and merges them as the baseline. It computes the Kolmogorov–Smirnov distance of the current window from that baseline: the CDF is rank/count, compared at both windows' percentiles. Above the threshold (0.1) the series has drifted. Drift start and end are OTLP log events to Loki, and asap_drift_ks and asap_drift are gauges. exact_answers runs the same test on the exact values. The scenario triples http.route-3's median at minute 10.

Result (56-core x86 server, 1-minute windows)

• Both methods flag only route-3, in the first window after the change: KS 0.505 (sketch) and 0.491 (exact), against a no-change noise level of 0.012–0.017 and 0.001–0.003. There were no false positives across the other 19 routes.
• The KS distance decays as the new distribution enters the baseline.
• The exact baseline's memory grows steadily because it must retain 5 windows of raw values; the sketch stages stay flat.

cargo fmt --check, cargo clippy --all-targets --features otap-engine,stream-benchmark -- -D warnings and cargo test --features otap-engine,stream-benchmark pass at this PR's head.