feat(closed_loop): shard evaluation across all DDP ranks

[HansRobo]

Problem

closed_loop_validate runs inside if global_rank == 0:, so one GPU does the evaluation while
the rest idle at the next collective. On a 4-GPU job that is 48 minutes of 3 idle GPUs per
checkpoint (102 minutes on the final save, which also renders media).

The DDP machinery for this already exists (shard_items, run_distributed, merge_ddp_shards)
but has no caller. It also barriers once per (site, object-mode), so a site with one route runs at
single-GPU speed regardless of GPU count — measured ceiling 1.96x on 4 GPUs.

Fix

run_evaluations_ddp() pools all (combo, route) pairs into one list, splits it once with
cost-weighted LPT, runs with no intermediate barriers, then barriers once and merges per combo.
closed_loop_validate moves out of the rank-0 guard.

Result

4xH100, real manifest, 38 segments:

	sequential	4 GPU
media OFF	2862 s	1023 s	2.80x
media ON	6121 s	2736 s	2.24x

Summed per-rank rollout time matches the sequential baselines to 0.04% / 0.1% — same workload,
negligible contention.

Output is identical. All 38 mp4s match, including 14 routes written concurrently
into one directory.

For reviewers

• elapsed_sec is now all-reduced per combo, preserving its per-site meaning. This touches an
  existing per-site W&B metric deliberately.
• run_distributed() now has zero callers (it had none before either). Deleting it is your call.
• 2.80x vs an ideal 4.00x is the cost proxy, not the mechanism: frame count orders routes well
  within a site, poorly across sites (0.026-0.079 s/frame). Feeding measured per-route seconds
  back would reach 4.00x; that needs rank-0 to plan and broadcast, so it is left as follow-up.

Tests

25 tests, where there were 0. The multi-rank ones spawn real gloo groups on CPU and drive the
production merge path. Failure propagation is mutation-verified.