fix(nft): fully align NFT with DiffusionNFT

[Rockdu]

Stack 2/9 of follow-ups to #263: main ← #265 ← this ← #267 ← #268 ← #269 ← #271 ← #272 ← #273 ← #274.

Aligns the NFT path with the reference implementation (NVlabs/DiffusionNFT).

What

• NFT trains a random subset of --diffusion-nft-timestep-fraction of the sigmas per sample, instead of always dropping the smallest sigma.
• NFT reads π_old from the EMA directly; --ref-mode now only picks the KL reference (none / lora_base / ref), and NFT supports --diffusion-kl-beta (kl_beta * mean((v_new - v_ref)^2) per pair, logged as nft_kl, renamed to kl_loss in refactor(nft): log the NFT KL term as kl_loss like Flow-GRPO #272). --loss-type nft requires --use-ema --rollout-weights ema.
• The EMA decay schedule follows the actor's optimizer step count (EMAOptimizer.step(optimizer_step)), as DiffusionNFT's return_decay(global_step) does; the separate update_count is gone.

Why

Each item was a line-level mismatch with the reference code: a fixed dropped sigma never trains, π_old and the KL reference were one role, and the EMA ramp counted train calls instead of optimizer steps.

Validation

• fix(nft): train a random subset of sigmas ... moves the NFT metrics; test_sd3_nft_pickscore_3xGPU and test_krea2_nft_ocr_2xGPU are re-recorded bitwise on that commit. Every other standard is unchanged.
• refactor(nft): train against EMA pi_old ... and fix(ema): follow the actor's optimizer step count ... are metric-neutral for these recipes: π_old is the same EMA, and one optimizer step per rollout makes the step count equal the old update count. CI checks this against the standards above.

Not in this PR

• The GRPO reward std keeps torch's default correction=1 (as miles, verl, TRL and verl-omni do; flow_grpo, DiffusionNFT and UniRL use ddof=0), since neither choice is more principled and switching would re-record every GRPO standard.

Files

• miles/ray/data_conversion_hub/nft.py: per-sample sigma subset.
• miles/backends/fsdp_utils/actor.py, loss_hub/{nft,flow_grpo,sft}.py: old_pred next to ref_pred; NFT KL term.
• miles/backends/fsdp_utils/ema.py: decay from the optimizer step.
• miles/utils/arguments.py, recipes, docs: flag semantics.
• tests: test_loss_hub_nft.py, test_reference_models.py, test_ema_optimizer.py, test_reference_arguments.py, GPU offload worker; two re-recorded NFT e2e standards.

Checklist

• pre-commit run --all-files passes — not run locally
• Added/updated tests for new behaviour
• pytest -x is green — CPU fsdp and argument tests pass locally; GPU tests run in CI
• If launch flags changed, python3 train.py --help still parses — not run (needs the GPU image)
• If a public flag was added, it appears in the CLI reference docs
• If an example was added, it has a real walkthrough — n/a

[zhihengy]

LGTM