feat: wire per-site closed-loop eval into the training loop (PR 6/6)

[kosuke55]

Summary

Final part of the stacked closed-loop eval series (stacked on #302). Wires PR1-5 into the training loop:

closed_loop_validate() (train.py) now runs one FullRouteClosedLoopEvaluation (PR2) per (site, object-mode) pair:

• --closed_loop_npz_root (single arbitrary path) and --closed_loop_sites_npz_root (auto-discovered multi-site root, PR3's site_discovery) fire independently — not if/else — so an ad-hoc override-scene test and a full per-site sweep can run in the same training call, each contributing its own rows to the combined episode table / cross-site aggregate (PR4's wandb_closed_loop).
• Each source picks its own object modes via --closed_loop_npz_object_modes (default objects-only — usually a single curated scene, where the traffic-reaction ablation doesn't apply) / --closed_loop_sites_object_modes (default objects+noobj — the full per-site objects-vs-no-objects comparison).
• --closed_loop_sites_npz_root may instead be a curated .json path-list manifest (PR3's discover_sites_from_json) instead of a sites-root directory, so training-time site closed-loop validation can use the same curated manifest as --closed_loop_npz_root's existing JSON-list support, instead of recursively evaluating every route under each site's split dir.

train_config.py/train_predictor.py/train_grpo_predictor.py gain the corresponding CLI/config fields (site roots, object modes, seg_len, the abort_*/unstick_* RolloutParams knobs, wandb video-pick/colormap-metrics/report-base-url). train_run.py forwards --closed_loop_sites_npz_root alongside the existing --closed_loop_npz_root.

Closes out the stacked series: rollout engine (#298) → evaluation framework (#299) → site discovery/colormap (#300) → W&B dashboard/HTML report (#301) → standalone CLI (#302) → this.

Test plan

• ast.parse + module import check + TrainConfig instantiation check
• Verified end-to-end on a real 8-GPU server training run (real dataset, real model): both sources firing together, both object modes, per-site + cross-site wandb aggregate, combined episode table, curated W&B dashboard
• Fixed a real crash surfaced during that run (stale s.n_snaps field reference in PR1's early-abort debug logging, now in PR1/feat: closed-loop rollout engine — objects/no-objects ablation + graded metrics (PR 1/6) #298)
• Verified --closed_loop_sites_npz_root as a .json manifest through the real, unmodified train_run.py CLI entrypoint (not just a direct function call): resumed a checkpoint, ran 10 real epochs, closed-loop fired at the checkpoint-save epoch against a real 2-site JSON manifest, all 4 (site, object-mode) combinations completed with correct metrics, videos and W&B log uploaded successfully