feat(otel): OTLP/HTTP trace ingress

[prashar32]

RiskKernel already emits OpenTelemetry GenAI spans to your backend (the export
side of Surface 3). This adds the other half — ingress: RiskKernel can act as
an OTLP/HTTP trace endpoint at POST /v1/traces and meter GenAI spans from
apps already instrumented (OpenLLMetry, the OpenAI Agents SDK, the Vercel AI SDK),
so spend is visible for agents that never routed through the proxy or SDK.

Point an existing app's OTLP exporter at the daemon and the model calls it already
traces show up against governed runs, with tokens and cost metered into the same
ledger the proxy uses — the usual one env var:

RISKKERNEL_OTEL_INGRESS_ENABLED=true riskkernel serve
# on the app:
export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:7070
export OTEL_EXPORTER_OTLP_HEADERS=authorization=Bearer $RISKKERNEL_API_TOKEN

What it does

• POST /v1/traces, the standard OTLP path, accepting both encodings —
  protobuf (application/x-protobuf, the SDK default) and JSON — and replying with
  an OTLP ExportTraceServiceResponse in the same encoding.
• Meters each span carrying gen_ai.usage.* token counts: reads the model
  (gen_ai.response.model, falling back to gen_ai.request.model) and
  gen_ai.system, prices via the existing table, and records through the ledger.
• Correlates to a run by riskkernel.run.id (on the span, falling back to the
  resource), creating the run lazily under the default budget like the proxy's
  run-id header. A GenAI span with no run id is observed and reported as a
  rejected span in the OTLP partial-success response. Spans without usage (tool
  calls, retrieval, framework spans) are ignored. Token counts emitted as a double
  or numeric string are accepted rather than dropped to zero.
• Off by default (RISKKERNEL_OTEL_INGRESS_ENABLED) and authenticated like the
  rest of the API.

Scope

Observe + meter: an ingested call already happened in the other app, so it's
recorded against the run's ledger (and marks the run halted if it crosses the
budget) but not blocked after the fact. Governing consumed spans — refusing the
next call on an externally-driven run — is a separate, future step. For
before-the-fact enforcement, route through the proxy (Surface 1) or SDK (Surface 2).

Dependency note

Uses the canonical OTLP proto types (go.opentelemetry.io/proto/otlp) and
google.golang.org/protobuf, both already in the module graph transitively via
the OTLP exporter we use for egress — promoted from indirect to direct, so this
adds no new module to go.sum. Hand-rolling protobuf parsing of
ExportTraceServiceRequest would be far more error-prone.

Tests

• internal/otel/ingress_test.go: protobuf and JSON over HTTP (verbatim decode +
  metering + the partial-success reply), run id from the span and from the
  resource, pricing of a consumed call, lenient numeric attribute types, and
  skipping of non-GenAI and unattributed spans.
• Verified live end-to-end: started the daemon with ingress enabled, POSTed an
  OTLP JSON span with riskkernel.run.id, and confirmed the run appeared with the
  metered tokens and a priced cost.
• go test -race ./... green; go vet ./... clean; gofmt clean.

Pinned attributes are documented in api/v1/otel-genai.md; user-facing usage in
docs/OTLP_INGRESS.md.

Closes #90