feat(telemetry): per-stage review spans for stage-level observability

[mountainowl]

Makes bubo emit richer observability data, not an observability tool — exactly the ask: bubo provides the metrics/traces; you bring Prometheus/Grafana/Tempo/Honeycomb/whatever.

What

Each review is one trace. llm_review.run now has a child span per stage, so a trace shows where the wall-clock actually went (spoiler: it's the agent, not bubo):

llm_review.run                  repo, mr_iid, sha, run_id
├─ llm_review.checkout          repo, sha
├─ llm_review.provenance        repo            (only when governance is on)
├─ llm_review.agent             repo, model, tokens_input/output/total, cost_usd, exit_code
└─ llm_review.post              repo, dry_run, findings_posted/planned/skipped

So any OTel backend can break latency, tokens, and cost down by stage and by model without bubo shipping a dashboard.

How

• ReviewTelemetry.span now uses start_as_current_span, so nested span() calls become real children (was start_span, which left them as siblings). Disabled-noop and exception-propagation behavior preserved.
• New set_span_attrs(span, **attrs) helper (mirrors add_event).
• The worker wraps each stage in a child span; the agent/post spans get token/cost/count attributes.
• No new dependency (existing OTel SDK), no-op when telemetry is off (the default), zero behavior change to the review itself.

Verify

416 tests pass, incl. a new in-memory-exporter test asserting the agent span is a child of the run span and carries its attributes. ruff + mypy + mkdocs --strict clean; cz valid; no footer. docs/telemetry.md gains a span-tree section.

Note: trivial CHANGELOG [Unreleased] overlap with the open docs PR #100 — keep both bullets.