docs(observability): add Grafana dashboard starter

[everettVT]

Summary

Adds an importable Grafana dashboard for Daft observability, built on the OTel metrics already documented in docs/observability/telemetry.md. 5 panels: rows/sec by operator type, bytes flow, task lifecycle, top operators by cumulative duration, and a failed-tasks stat panel.

Companion to the in-process Daft Dashboard (daft dashboard start):

• In-process dashboard → live per-query debugging (plan tree, tasks view, heatmap, event log replay)
• Grafana dashboard → fleet monitoring / on-call / SLO tracking alongside your existing infra observability

Why a draft

Shipped untested end-to-end. I built this from the documented metric surface but haven't run it against a live Daft + OTel + Prometheus + Grafana stack. The metric query shapes assume the standard OTel Collector → Prometheus exporter naming convention (daft.X.Y → daft_x_y_total); this should be verified against an actual scrape before users follow the docs.

Files

• docs/observability/grafana.md — new docs page (prerequisites, import instructions, panel reference, metric naming convention, "what's not covered yet", adaptation tips)
• docs/observability/grafana/daft-dashboard.json — the dashboard JSON (Grafana 10+, schema v39, $DS_PROMETHEUS datasource variable)
• docs/SUMMARY.md — adds "Grafana Dashboard" entry under Observability nav, after Telemetry

Test plan

• Import daft-dashboard.json into a real Grafana instance pointed at a Prometheus scraping OTel-exported Daft metrics
• @samstokes / @cckellogg / @universalmind303 — verify metric query names against actual Prometheus output (daft_rows_out_total, daft_duration_total, daft_task_*, etc.)
• @desmondcheongzx — confirm the process-level memory + CPU metrics (PR feat(observability): add process-level memory and CPU monitoring #6428) are visible via Prometheus and what their names are, so a future PR can add a "process resources" panel row
• Verify Top 10 operators by cumulative duration query (topk(10, sum by (node_type, node_id) (daft_duration_total))) returns expected shape — the µs counter may need to be rescaled in the panel display
• Confirm the "distributed-only" caveat in the docs page matches reality (we list daft_bytes_read_total, daft_task_active, daft_task_* as distributed-only per telemetry.md)
• When @samstokes' peak resident state PR (feat(metrics): track peak resident state bytes for stateful operators #6883) lands, add a memory-attribution panel
• Once per-operator bytes_in / bytes_out Prometheus exposure is confirmed, add inflation/deflation panels

Context

This started as a marketing-repo canon artifact (Eventual-Inc/marketing#360) supporting the W21 observability launch. Moved here because:

1. Daft repo can actually test it end-to-end against a real stack
2. Users reading docs.daft.ai/observability/ will find it where they're already looking
3. Engineering reviews the metric queries against actual scraper output before users import it

The marketing PR (#360 in Eventual-Inc/marketing) will be closed in favor of this one.

[github-actions]

Rust Dependency Diff

Head: 2407e359aa6076108ab9153a0503b54365846539 vs Base: 4dd8ee211db381293f8ac75c9e47c9c160d560ce.

✅ OK: Within budget.

• New Crates: 0
• Removed Crates: 0