What does this PR do?

Type of change: new feature

Adds warm-start fine-tuning for DFlash: a new dflash_init_checkpoint field on DFlashConfig loads an already-trained, exported drafter (DFlashExporter / z-lab layout) into model.dflash_module at conversion time, so training continues from it instead of random init. Must be a local directory containing model.safetensors (stage public drafters with hf download first).

• Shape/key mismatch vs dflash_architecture_config fails with a clear error; mask_token_id / target_layer_ids mismatch vs the checkpoint's config.json logs a degradation warning. dflash_block_size may differ (weights are block-size agnostic).
• Domino/DSpark can warm-start their DFlash backbone from a plain DFlash checkpoint (missing head submodules train from scratch, with a warning).
• Restore of a saved training checkpoint ignores the field (restore_dflash_model clears it), so resume works even if the init checkpoint is gone.

Also adds a launcher example (tools/launcher/examples/Qwen/Qwen3-8B/hf_online_dflash_finetune.yaml) and a doc section in examples/speculative_decoding/doc/dflash.md.

Usage

# recipe YAML (or launcher dotlist: dflash.dflash_init_checkpoint=...)
dflash:
  dflash_init_checkpoint: /hf-local/z-lab/Qwen3-8B-DFlash-b16  # local dir with model.safetensors
  dflash_block_size: 16
  dflash_mask_token_id: 151669
  # must match the checkpoint's draft architecture (dims do NOT inherit from the base model):
  dflash_architecture_config:
    num_hidden_layers: 5
    num_attention_heads: 32
    num_key_value_heads: 8
    head_dim: 128
    intermediate_size: 12288

uv run launch.py --yaml examples/Qwen/Qwen3-8B/hf_online_dflash_finetune.yaml --yes

Testing

• New unit tests: TestDFlashWarmStart (export→load round-trip at a different block size, missing-dir / arch-mismatch errors, restore ignores the field). All dflash/domino/dspark unit tests pass.
• End-to-end: converted Qwen/Qwen3-8B with dflash_init_checkpoint pointing at a local copy of z-lab/Qwen3-8B-DFlash-b16 — all 58 tensors load bit-exact, target_layer_ids=[1, 9, 17, 25, 33] matches the checkpoint, no warnings.

Before your PR is "Ready for review"

Make sure you read and follow Contributor guidelines and your commits are signed (git commit -s -S).

Make sure you read and follow the Security Best Practices (e.g. avoiding hardcoded trust_remote_code=True, torch.load(..., weights_only=False), pickle, etc.).

• Is this change backward compatible?: ✅ (new optional field, default None; existing configs/checkpoints unaffected)
• If you copied code from any other sources or added a new PIP dependency, did you follow guidance in CONTRIBUTING.md: N/A
• Did you write any new necessary tests?: ✅
• Did you update Changelog?: ❌ (can add if desired)
• Did you get Claude approval on this PR?: ❌ (pending)