feat(tracing): provider-agnostic exporters with first-class PostHog support

[Spherrrical]

Summary

Adds a provider-agnostic telemetry export layer to brightstaff, driven by a new tracing.exporters config list, and ships PostHog as the first first-class integration. Users point at their PostHog URL + project token and every LLM call is streamed as a $ai_generation event — as simple as LiteLLM's posthog callback.

tracing:
  random_sampling: 100
  exporters:
    - type: posthog
      url: https://us.i.posthog.com
      api_key: $POSTHOG_API_KEY
      distinct_id_header: x-user-id   # optional; else anonymous

Changes

• Config: new tracing.exporters[] (schema + Exporter/PosthogExporter in common), extensible to other backends via an internally-tagged enum.
• PostHog exporter: new SpanExporter (posthog_exporter.rs) that filters LLM spans, maps attributes to $ai_* properties (model, provider, latency, tokens, http status, trace ids), and POSTs a /batch/ payload via reqwest. Failures are logged and dropped, never blocking requests.
• Wiring: init_tracer now composes multiple batch exporters (OTLP + each PostHog exporter); enable gate decoupled so PostHog works without opentracing_grpc_endpoint.
• distinct_id: resolved from a configurable request header into AppState, stamped on the span as plano.distinct_id; anonymous fallback (\$process_person_profile=false).
• Span enrichment: now records llm.provider. capture_messages opt-in sends the truncated user preview as $ai_input.
• Docs/tests: tracing guide + full reference config updated; Rust config + span-mapping tests and a CLI schema test added.

Note: stacked on musa/agent-skills-support (PR #954) to keep this diff scoped to the PostHog work.