gooddata-eval: fix dropped checks and over-broad exception catches from PR #1845 review #1942

Three CodeRabbit findings on #1845, not addressed before merge (non-blocking, tracked here instead):

1. packages/gooddata-eval/src/gooddata_eval/core/agentic/alert_skill.py:1266-1284 — feat(gooddata-eval): populate best_run_latency_s for agentic evaluators #1845 (comment)
   The best-run detail now hardcodes 9 fixed checks instead of **ev.checks(). Fixture-conditional checks (comparison_correct, title_correct, dashboard_correct) are silently dropped, defaulting to True in the report and inflating quality_score. The per-run _run_detail still uses checks(), so winning and failed runs now have different keys. Fix: use **ev.checks() and include the conditional checks in the failure message too.

2. packages/gooddata-eval/src/gooddata_eval/core/agentic/anomaly_detection.py:415-424 — feat(gooddata-eval): populate best_run_latency_s for agentic evaluators #1845 (comment)
   except Exception around send_message labels any failure LoopExit.CHAT_ERROR, including non-chat bugs (e.g. a TypeError in _accumulate), mislabeling them as a GoodData infra fault. Should catch only ChatError (and httpx.HTTPError if transport faults must be recorded), matching kda_skill.py's existing pattern.

3. packages/gooddata-eval/src/gooddata_eval/core/agentic/what_if.py:372-381 — feat(gooddata-eval): populate best_run_latency_s for agentic evaluators #1845 (comment)
   Same over-broad except Exception → CHAT_ERROR pattern as (2).