perf(minimax-h3): batch replay timesteps

[leviking98z-rgb]

Summary

• add an opt-in batch_replay_steps path for MiniMax-H3 that evaluates multiple replay timesteps in one step-major transformer batch
• preserve H3's separate video/audio sigma schedules and existing element-weighted joint A/V log-prob
• extend packed timestep metadata and the vendored transformer AdaLN lookup to accept per-batch row timestep indices without concatenating attention documents

Related Issue

• Research plan and exactness contract: [Research][Systems] MiniMax-H3 视频后训练加速：前沿 MLSys 技术映射与实验计划 leviking98z-rgb/UniRL#42
• Related generic/video-model implementation: feat(diffusion): batch video replay timesteps #394

Test Plan

Not run; reason: code-only Draft PR requested. GPU forward/backward parity and MiniMax-H3 end-to-end timing remain pending.

Compatibility / Risk

• batch_replay_steps defaults to false, so existing recipes keep the serial replay path.
• No checkpoint or data-format changes.
• The new path increases replay activation memory by stacking selected timesteps on the batch axis; the practical K2/K3 limit still needs GPU measurement.
• Batched replay uses the repository's existing replay-anchor safety check, so configurations must use old_logp_source: replay when enabling it.

Reviewer Notes

• This PR is independent of feat(diffusion): batch video replay timesteps #394 because the shared replay helper and replay-anchor guard are already present on main; feat(diffusion): batch video replay timesteps #394 does not cover MiniMax-H3's dual trajectory and per-modality sigma grids.
• The implementation keeps each timestep as a separate batch item and therefore a separate attention document. Only the row-to-AdaLN timestep mapping becomes batch-aware.
• AI assistance: implemented with Codex. Duplicate-work check covered open PRs/issues matching MiniMax-H3, replay, and batching; no open PR currently adds H3 replay timestep batching.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.