refactor(diffusion): extract loss_hub for Flow-GRPO and custom loss dispatch

[niehen6174]

Motivation

Move Flow-GRPO PPO-clip and GRPO reward normalization out of the actor/rollout into swappable building blocks — same default behavior, ready for NFT/SFT as flag recipes.

• Problem: ~300 lines of Flow-GRPO math lived inside actor.py; reward normalization was duplicated inline in rollout.py. There was no working hook to replace the train objective.
• Fix: New loss_hub/ module + --loss-type flow_grpo / --custom-loss-function-path. Actor only schedules micro-batches and calls loss_fn(ctx, batch).
• Scope: Structural refactor only. Default Flow-GRPO path is a move, not a rewrite. NFT / SFT / offline collection are follow-up PRs.

---

Changes

BEFORE                              AFTER
────────────────────────────────    ────────────────────────────────────
RolloutManager                      RolloutManager
  └─ GRPO norm (inline, 25 lines)     └─ grpo_normalize_rewards()  ← loss_hub
  └─ RolloutTrainDataConverter        └─ RolloutTrainDataConverter (unchanged)

FSDPTrainRayActor (~300 lines)      FSDPTrainRayActor (~20 lines of loss glue)
  └─ DiT forward                    get_diffusion_loss_function(args)
  └─ SDE logπ                       └─ flow_grpo_ppo_loss(ctx, batch)
  └─ PPO-clip
  └─ micro-batch schedule             micro-batch schedule (unchanged)
                                      backward / optim (unchanged)

Roadmap

flowchart LR
  P1["PR1: loss_hub<br/>(this)"] --> P2["PR2: NFT<br/>swap convert + loss + signals"]
  P2 --> P3["PR3: SFT<br/>swap rollout + convert + loss"]

  style P1 fill:#E8F1FF,stroke:#2F6FED
  style P2 fill:#EAF8EF,stroke:#1B7F4E
  style P3 fill:#FFF6E5,stroke:#C47D00

Loading

Follow-up	Hooks to swap	Notes
DiffusionNFT	--custom-reward-post-process-path, --custom-convert-samples-to-train-data-path, --custom-loss-function-path	Forward FM; no traj/logπ needed
SFT	--rollout-function-path, convert, loss, --disable-compute-advantages-and-returns	Same pattern as Miles sft_rollout + sft_loss
Collection tuning	rollout params	Skip unused traj/logπ once NFT lands

---