docs(examples/monitoring): add chatd lifecycle Grafana dashboard - #28975

Grafana dashboard for the chatd lifecycle metrics at examples/monitoring/dashboards/grafana/chatd-lifecycle/, with a README documenting every panel.

Rows, sharing one dashboard-level $stat selector (mean, p50, p90, p95, p99): a stat panel showing the --chat-stage-metrics level from coderd_chatd_stage_metrics_level; the turn time partition (per-model mix, seconds per turn by category, unattributed time), read from the turn_time_seconds_total and turns_total counters so it works at basic; one row per stage tree level with duration per occurrence for the stages observed at the default basic level; throughput, TTFT, and background provider calls. Each level also has a collapsed row titled (full level only: --chat-stage-metrics=full) holding the per-turn panels (seconds per turn, occurrences per turn, share of turn), the per-turn category share, and the stages observed only at full (generation_step, prepare, thinking, compaction). The label schema is the same at every level, so those panels read "No data" rather than erroring at basic. Variables: $datasource, $stat, $chat_kind, $organization, $model; $organization is applied wherever $chat_kind is. Every division guards its denominator and every histogram_quantile keeps le.

The README documents the levels with an annotated stage tree, the series cost per level, example alert rules on the round bucket edges, and a scrape-time relabel snippet for alert-only deployments.

An earlier revision had an aggregate flamegraph and hierarchy table row built from a 15-stage label_replace expression over [$__range]. It was removed: the query loaded every bucket series for the whole dashboard range on each refresh, and because Prometheus cannot compute self time for overlapping stages the flamegraph shape was misleading. The per-level rows carry the same per-stage data with bounded lookback.

Verified against a live dev deployment (Grafana 12.1, Prometheus, Tempo) at both basic and full: every expanded panel populates at basic with 148 series after three root turns on one model (326 before the 12-bucket ladder and the counter-based partition), and the collapsed rows populate at full (1,132 series after six turns).

Part of the chatd lifecycle observability stack: 1 extract step · 2 stage tracer · 3 stream stages · 4 provider attempts · 5 turn span · 6 capacity wait · 7 turn accounting · 8 dashboard. Supersedes the experimental branch in #28741.

🤖 This PR was generated by Coder Agents on behalf of @jscottmiller.