Refresh training dashboard semantics

[k15z]

Summary

• make the dashboard mode-aware so approved-baseline views do not surface stale candidate gate artifacts
• add trainer live/resume metadata and config thresholds to the status API
• replace imprecise policy+value loss chart with all-head/total trainer loss and add cadence/self-play breakdown views
• tighten benchmark, version-history, and recent-game tables for desktop and mobile scanning

Verification

• uv run pytest training/test_dashboard_store.py training/test_trainer_artifacts.py training/test_trainer_promotion.py
• make lint
• codex review --uncommitted
• Playwright desktop/mobile screenshots against local dashboard on real S3 data; browser console clean