feat(autoresearch): prompt optimizer + API endpoints + M3 integration (#3200)

[mrveiss]

Summary

• PromptOptimizer with LLM mutation → benchmark → scorer chain → keep/discard loop
• PromptOptTarget, PromptVariant, OptimizationSession models
• 10 new API endpoints: prompt optimizer (start/status/cancel/variants/score), approvals (list/decide), insights (list/search/synthesize)
• Notification dispatch on approval request via NotificationService
• Updated __init__.py with all M3 exports
• Integration test: optimize → synthesize → query insights

Depends on: #3202 (Scorers) and #3203 (Knowledge Synthesizer) — both branches merged into this one

Closes #3200
Part of #2600 (AutoResearch M3)

Test plan

• 16/16 tests pass (4 optimizer + 11 scorers + 1 integration)
• Pre-existing test failures unchanged

🤖 Generated with Claude Code