feat: support offline DFlash-family training

[maocheng23]

Motivation

#678 adds unified-runtime offline consumers, but the production feature preparer is still EAGLE3-specific: it requires exactly three layers and writes hidden_state/aux_hidden_state. DFlash and Domino consume hidden_states, while DSpark additionally needs target_last_hidden_states.

This change makes the producer schema come from the same typed algorithm registry as the consumer, so offline preparation and training cannot silently drift.

Modifications

• add algorithm-owned offline capture layouts describing the SGLang hook, auxiliary/final-state mapping, and passthrough tensors
• generalize scripts/prepare_hidden_states.py for strategy-aware draft configuration, arbitrary layer counts, and exact persisted schemas
• support DFlash and Domino feature production, plus DSpark offline storage, normalization, collation, and runtime wiring
• dispatch through set_dflash_layers_to_capture for the DFlash family and set_eagle3_layers_to_capture for EAGLE3
• add offline recipes for Qwen3 DFlash, Domino, and DSpark, with capability and data-preparation documentation

D-PACE and DTA inherit DFlash's capture/storage contract. P-EAGLE is intentionally unchanged because its target-head/tokenizer preparation is a separate algorithm path.

Related Issues

Follow-up to #678.

Accuracy Test

• exact-commit DSpark CUDA/FSDP gate on H200: passed two training steps with finite parameters
• exact-commit two-H200 DFlash capture on pinned SGLang 0.5.14: passed on both ranks with five captured layers and the expected 5x auxiliary width

Benchmark & Profiling

Not applicable. This changes offline data preparation and provider wiring; it does not change kernels, model forward math, or serving performance.

Validation

• pre-commit run --all-files — passed
• focused provider/composition/capture/config/architecture suite — 79 passed
• adjacent config and offline-runtime suite — 58 passed, 3 CUDA-only skips on macOS
• DSpark H200 runtime gate — passed
• pinned SGLang 0.5.14 two-H200 DFlash capture gate — passed

Stack

This draft is stacked on #678 at its current head, ffe081c8074eaca7a079889737f82cb6649569de. Until #678 merges, GitHub will display its commits in this draft too; afterward the diff reduces to this offline DFlash-family follow-up.

Checklist

• Format the code with repository pre-commit hooks.
• Add unit and runtime tests.
• Update documentation, recipes, and docstrings as needed.
• Provide accuracy/performance validation as applicable.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.