AutoResearch: wire up PromptOptimizer API with real scorers and benchmark

[mrveiss]

Summary

The prompt optimizer API endpoint and singleton have placeholder/empty wiring that makes the API non-functional:

1. Empty scorers dict — `_get_optimizer` initializes with `scorers={}`, so any scorer chain reference (e.g., `"val_bpb"`) hits "scorer not found, skipping" and produces unscored variants
2. Placeholder benchmark — `start_optimization` endpoint uses `async def _benchmark(prompt): return prompt` which just echoes prompts back, producing meaningless optimization results
3. No agent registration path — the `PromptOptTarget` is hardcoded for `autoresearch_hypothesis` only, with no mechanism for other agents to register

Expected Behavior

• Default scorers (ValBpbScorer, LLMJudgeScorer, HumanReviewScorer) registered at optimizer init
• AutoResearchAgent's hypothesis generation wired as the real benchmark function
• API start endpoint validates target exists and uses real benchmark

Files

• `autobot-backend/services/autoresearch/routes.py` (lines 240, 306-307)

Origin

Discovered during code review of PR #3206 (Issue #3200)