Add importable Datadog dashboard example

[prashar32]

Completes the OTel dashboard set — Grafana + Tempo and SigNoz already shipped; this adds the Datadog import so a team already running Datadog can visualize governed runs straight from the OpenTelemetry GenAI spans RiskKernel exports. No extra instrumentation: Datadog ingests the spans over OTLP, and you import one JSON file.

What's here

• examples/otel/datadog/dashboards/riskkernel-agent-runs.json — Datadog's importable dashboard schema (layout_type: ordered, data_source: spans). Panel parity with Grafana/SigNoz:
  • Total spend, output tokens, model calls, tool calls refused (query-value tiles)
  • Spend over time, spend by run (toplist), p95 latency by model, model-call rate by model, output tokens by model, tool calls by outcome, budget halts by reason
• examples/otel/datadog/README.md — how to point the OTLP exporter at the Datadog Agent's OTLP receiver (gRPC 4317 / HTTP 4318), import the dashboard, and what each panel queries.

Datadog specifics (called out honestly in the README)

• The Datadog key lives on the Agent, not RiskKernel. RiskKernel only emits OTLP to the endpoint you give it (no phone-home); DD_API_KEY is standard Agent config.
• Span attributes are queried with the @ prefix (@riskkernel.cost.usd, @gen_ai.request.model); tags have none. Datadog's OTLP intake preserves the dotted names, so they read identically to the pinned set.
• @duration is Datadog's intrinsic span-duration measure (nanoseconds) for the latency panel — intrinsic, not a RiskKernel attribute.
• Indexed spans + retention. data_source: spans panels query retained spans; the README notes keeping a retention filter for RiskKernel's spans, and documents the optional span-based metric upgrade for durable, long-window cost rollups (and the billing note).
• Service name maps to Datadog's service tag; the dashboard ships a $service template variable defaulting to riskkernel.

Contract fidelity

Every attribute the queries touch matches api/v1/otel-genai.md exactly (verified programmatically): gen_ai.operation.name, gen_ai.request.model, gen_ai.usage.output_tokens, riskkernel.cost.usd, riskkernel.run.id, riskkernel.tool.status, riskkernel.halt.reason — plus the built-in @duration. The JSON parses cleanly. Cross-linked from the Grafana, SigNoz, and main observability READMEs, and a CHANGELOG entry under [Unreleased].

Closes #89