[Metrics] Query latency histogram

[alexkarp-umd]

PgDog exposes query latency as total_query_time and avg_query_time. Both are means over a pool. As an on-call SRE I want to have percentiles available since averages alone tell you very little about worst-case performance about the distribution of request latencies.

I propose adding a per-pool histogram of individual query durations to OpenMetrics and OTLP, so histogram_quantile() works against PgDog's metrics:

histogram_quantile(0.99, sum by (le, database) (rate(pgdog_query_time_seconds_bucket[5m])))

Proposal

• OpenMetrics: _bucket / _sum / _count, cumulative counts, trailing le="+Inf".
• OTLP: a real Histogram data point with explicitBounds / bucketCounts and support for both cumulative and delta mode (see [Metrics] Datapoints follow OTEL_EXPORTER_OTLP_METRICS_TEMPORALITY_PREFERENCE #1293)
• Default ladder 100µs → 30s in 12 bounds, plus the implicit +Inf, so 13 series per pool.
• Bucket count and boundaries configurable at startup.

Status

I have this working on a branch rebased onto #1293 and I can open a PR for it.