feat: roll out DiffusionNFT under the EMA shadow adapter

[Jayce-Ping]

Summary

Roll out DiffusionNFT under its EMA-smoothed old shadow adapter on a SEPARATE rollout engine (SGLang / vLLM-omni), the way requires_ema_rollout intends. The in-process apply_eval_ema swap only touches the train process, so a separate engine never saw the EMA weights.

Reworked (rebased fresh on latest main) into a single, transport-agnostic mechanism owned by the backend:

• Backend single source of truth (unirl/train/backend/fsdp.py): FSDPBackend.rollout_adapter_name returns the EMA shadow ("old") for adapter-EMA, else "default". The in-process eval-EMA swap and the cross-process weight sync both derive from it, so they cannot disagree.
• One unified adapter_name knob across both weight-sync families (weight_sync/full/*, weight_sync/lora/*): defaults to None -> asks the backend; an explicit value overrides. Full-sync folds the shadow into the merged base (SGLang); LoRA-sync ships the shadow adapter directly (vLLM-omni). Also fixes a latent bug where _extract_canonical_lora never forwarded adapter_name, so the LoRA path always shipped default.
• Trainer untouched: no per-recipe config-reaching; the sync derives the adapter from the backend.
• FSDP grad fix (unirl/train/inject.py): _activate_keep_grad re-asserts the trainable/frozen requires_grad split across rollout/restore swaps, so the trainable default is never all-gathered grad-less under reshard_after_forward=False.
• SGLang ODE path (unirl/rollout/engine/sglang/*): DiffusionNFT's num_sde_steps=0 resolves sde_indices to []; request the deterministic ODE rollout, always collect the dit-trajectory, and keep the full trajectory.

Related Issue

N/A

Test Plan

• python -m py_compile on all touched files: passes.
• Linter on touched files: clean.
• Standalone CPU logic checks (no torch in env): backend rollout_adapter_name (old with ema_lora_cfg, else default); both sync families resolve None -> backend; full-sync fail-closed guard fires for non-default + lora_merged=False; _extract_canonical_lora forwards adapter_name; SGLang empty-sde_indices -> ODE branch + full trajectory; EMA update math.
• Not run; reason: no local GPU. Recommended e2e: examples/diffusion/sd3_nft.yaml (trainside trains past the first backward) and examples/diffusion/sd3_nft_sglang.yaml (separate, rolls out under EMA old).

Compatibility / Risk

• Backward compatible: GRPO and plain-LoRA recipes resolve adapter_name to "default" (unchanged behavior). FullWeightSync gains an adapter_name ctor arg (defaults to None); the LoRA-sync adapter_name default changes from "default" to None (resolves to "default" for non-EMA backends), so existing recipes are unaffected.
• Requires the sync handler's backend to expose rollout_adapter_name; FSDPBackend is the only training backend, so all paths are covered.

Reviewer Notes

• AI-assisted (Cursor). Rebased fresh on latest main; this force-push supersedes the earlier full-sync-only lora_adapter + trainer _resolve_sync_lora_adapter approach with the backend-sourced unified mechanism.
• Out of scope: no new vLLM-omni NFT recipe is shipped, but the unified design makes it work via either RemoteLoraWeightSync (ships old, no merge) or IPC/NCCLWeightSync + lora_merged: true (merged old base).

Checklist

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

[celve]

Do we have the e2e test result?