fix: share forward pass / loss in DFlash / DSpark

[Dogacel]

Overview

1. Share backbone forward pass in DSpark / DFlash.
2. Add L1 loss support (1/2 * TV) to DFlash training.
3. Update HF exporter to export vLLM / SGLang compatible keys directly

L1 loss (Same as TV) 0.5, CE loss 0.5 test run on 500 samples:

Also trained some internal checkpoints using 0.5 L1 / 0.5 CE loss and they slightly outperform CE only.

For HF exporting, I have re-exported existing DSpark endpoint (huggingface.co/Dogacel/Qwen3-8B-DSpark) and run,

vllm serve Qwen/Qwen3-8B --speculative_config '{"method":"dspark","model":"deepseek-ai/dspark_qwen3_8b_block7","num_speculative_tokens":7, "attention_backend":"FLASH_ATTN", "draft_sample_method": "probabilistic"}'

On the vllm branch: (https://github.com/Dogacel/vllm/tree/dspark-local).

It got ~2.3 acceptance length on SPEED-bench.

Future Work: Potentially moving D-PACE in DSpark's CE objective to get better acceptance, however the anchoring offset is defined differently in DSpark and DFlash, need to look closer.