feat(diffusion): add SFT loss hub and pre-encoded data manager

[zhihengy]

What

Diffusion SFT on (video, prompt) datasets, reusing the PR #63 hook architecture end to end. No sglang engines are involved: --loss-type sft_loss flips on the existing debug_train_only wiring, so placement groups, weight sync, and the driver loop need no changes.

How

• loss_hub/sft.py: prepare_sft_batch samples grid sigmas (phase-pure per micro-batch for dual-expert Wan2.2 via component_for_timestep), corrupts cached clean latents with x_t = (1-σ)x₀ + σε, and collates cached cond through the family TrainPipelineConfig; sft_loss_formula is velocity MSE against ε - x₀. Both plug into the existing --custom-prepare-train-batch-path / --custom-loss-function-path slots (auto-filled by --loss-type sft_loss).
• miles/ray/sft_data_manager.py: replaces RolloutManager behind the same driver interface. generate(rollout_id) is stateless — per-epoch seeded shuffle, contiguous slice, TrainDataDPSplitter, and a flow-shifted training sigma grid emitted as scheduler meta — so resume needs no cursor state.
• scripts/sft_encode_wan.py: one-time Ray encode job (UMT5 + Wan VAE from diffusers) producing one .pt per sample with the pair schema the prepare hook consumes.
• scripts/run-diffusion-sft-wan22.sh: 4-GPU Wan2.2 dual-expert LoRA SFT recipe.
• arguments: --sft-data-path; sft_loss validation rejects RL-only combos (--eval-interval, --diffusion-kl-beta, --diffusion-recompute-old-log-prob, --ref-mode, --ema-shadow, --n-samples-per-prompt != 1) and requires --diffusion-flow-shift for the training grid.

The rollout-side data_conversion_hub (expand hook) is intentionally not used: its (samples, rewards) signature is rollout-specific and SFT has neither.

Validation

• tests/fast/backends/fsdp_utils/test_loss_hub_sft.py (5 tests, pass): corruption/target identity, timestep scaling, dual-expert phase purity, loss values.
• SftDataManager exercised directly: shift math (σ(0.5) → 0.7507 at shift 3.0), terminal sigma, DP shard shapes, exact one-epoch coverage, different shuffle across epochs.
• Full parse_args round-trip for the recipe arg set (hooks auto-filled, debug_train_only=True, gbs derivation) plus rejection of each invalid combo.
• Not yet run on GPU — draft for design review first.