Add plan mode + structured outputs to review pipeline

- New Phase 1: _create_review_plan() uses agentic loop with tools to
  investigate the PR, then structures findings into ReviewPlan via
  structured output (summary, risk_areas, investigation_findings)
- Replace _gather_strategic_context with _plan_and_execute_searches
  that takes the ReviewPlan as context for better-targeted searches (3-6)
- All review passes (_pass1/2/3, _quick_wins, _llm_aggregate, judge)
  now use structured output (ReviewResult/AggregatedResult schemas)
  instead of fragile JSON parsing
- Add plan context to all review pass prompts via plan parameter
- Delete _parse_issues (45-line text.find("[") hack) - replaced by
  _parse_structured_issues using Pydantic validation
- Add Pydantic models: ReviewPlan, ReviewIssueSchema, ReviewResult,
  AggregatedResult