feat(flowgrpo): add bounded-memory chunked replay backward

[Yuanjia-Lee]

Summary

FlowGRPO replays all S SDE steps and runs one backward over the stacked
[B, S] log-probs, so every step's replay graph stays resident until that
backward. sd3_trainside.yaml already runs micro_batch_size: 1 with
activation_checkpointing: false; the resident graph still grows with S.

timestep_chunk_size makes FlowGRPO replay and backward SDE steps in slices
of that size, each scaled by len(chunk) / S; the summed gradient matches the
default single backward (DiffusionNFT's iter_scale pattern, optimizer
boundary left with TrainStack). Default None keeps the current behavior.

A/B on one rollout and anchor (S=4, micro 1), real FlowGRPO and
SD3DiffusionStage.replay: fp32 gradients match to 7.9e-8 relative with
identical loss and ratio metrics, and the CUDA peak in the compute drops
43.8 → 36.3 MB (chunk 1) and 40.8 MB (chunk 2): a config that cannot hold S
resident graphs can still train all S steps.

Related Issue

N/A

Test Plan

• SKIP=no-commit-to-branch pre-commit run --files unirl/algorithms/flowgrpo.py unirl/algorithms/README.md --show-diff-on-failure
  → all hooks pass.
• Harness on one 4090 (torch 2.11.0+cu130): real FlowGRPO over a real
  SD3DiffusionStage rollout, arms timestep_chunk_size None/1/2 on the
  same segment, beta 0 and 0.1. Numbers in the Summary.

Compatibility / Risk

• Default off; recipes unchanged. One extra reduce-scatter per chunk; bf16
  autocast shifts the gradient by 0.46% and the AdamW delta by 4.7e-6 at
  lr 3e-4 (fp32 matches).

Reviewer Notes

• Duplicate check: no open PR changes the FlowGRPO backward path. feat(diffusion): batch video replay timesteps #394 batches
  replay timesteps into one forward for throughput; no overlap.

Checklist

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