refactor(diffusion): require rollout scheduler sigmas, drop timesteps-derived fallbacks

[Rockdu]

Re-lands #89: it was accidentally merged into the already-merged #88 branch instead of main (the stacked base was not auto-retargeted because the branch was not deleted). Same three commits, rebased onto current main.

What

• Training now consumes the rollout engine's scheduler.sigmas snapshot verbatim; the timesteps / num_train_timesteps reconstruction fallbacks are removed (scheduler_meta_from_rollout, NFT converter, Flow-GRPO converter). Missing sigmas is a hard ValueError, making sglang 585a7d05e (#32683) the minimum rollout engine.
• Both converters now share scheduler_meta_from_samples, which builds the batch scheduler meta from sample 0 and rejects any sample whose timesteps/sigmas differ — the NFT converter previously used sample 0's meta without checking the rest.

Why

The fallback was not equivalent to the snapshot: traj.timesteps is [T+1] (terminal 0.0 appended with the final x0 latent), so cat([timesteps / N, zeros(1)]) produced 12 entries for a 10-step schedule vs the snapshot's 11, and the division does not round-trip sigma * 1000 in fp32 (rel. diff up to 1.42e-8, measured in CI run 30778705251). Two paths producing almost the same array is exactly how stage-c-3-gpu-h200 silently drifted when upstream started supplying sigmas; failing loudly removes that bug class.

Validation

• Numerical no-op by construction: sglang has populated dit_trajectory.sigmas unconditionally since 585a7d05e, so the deleted branches were already dead. Prior CI controls agree: pinning pre-sigmas sglang 3312645a3 reproduces the old standard 8/8 (run 30770528080), and nulling dit_trajectory.sigmas on current sglang reproduces it 8/8 (run 30772902308). No e2e standard re-record needed.
• Smoke test on 8x H200 (sglang main b8109b5d + this branch): all 5 recipes ran 2 rollouts end-to-end — SD3-OCR GRPO, SD3 NFT (nft_num_timesteps=9, nft_t_mean=0.7304325, matching the sigmas-snapshot theory value), LTX-2.3 GRPO, Qwen-Image Flow-GRPO-aligned, Wan2.2-A14B (dual-DiT per-component metrics present). Healthy log_prob_mean_abs_diff on all: ~1e-4 (SD3), ~1e-6 (LTX), ~4e-5 (Qwen-Image), ~1e-5 (Wan2.2).
• tests/fast locally: 136 passed (excluding test_metric_buffer_dist / test_hybrid_shard_mesh, which need a Linux dist/GPU environment).

Not in this PR (tracked separately): value-threshold trim for resolve_nft_sigmas (requires re-recording the NFT e2e standard), the degenerate terminal pair in next_timesteps, and the upstream [T] annotation fix (sgl-project/sglang#33332).

Files

• miles/utils/train_data_utils.py — scheduler_meta_from_rollout requires scheduler_sigmas; new shared scheduler_meta_from_samples.
• miles/backends/fsdp_utils/actor.py — drop the now-unused num_train_timesteps plumbing.
• miles/ray/data_conversion_hub/flow_grpo.py, miles/ray/data_conversion_hub/nft.py — use the shared helper; fallbacks removed.
• miles/utils/types.py — DiTTrajectory.sigmas comment updated.
• tests/fast/utils/test_grouping_parity.py, tests/fast/backends/fsdp_utils/test_loss_hub_nft.py — sigmas-required and mismatched-meta tests (registered stage-a-cpu).

Checklist

• pre-commit run --all-files passes — not run locally
• Added/updated tests for new behaviour
• pytest -x is green — tests/fast locally, see Validation
• If launch flags changed, python3 train.py --help still parses — N/A, no flag changes
• If a public flag was added, it appears in the CLI reference docs — N/A
• If an example was added, it has a real walkthrough — N/A