Qwen3-8B DSpark speculator

DSpark (DFlash backbone + Markov logit-bias head + confidence head) draft model for Qwen/Qwen3-8B, trained with speculators (PR #677).

Preview/research checkpoint: online training on a 20k ShareGPT slice, seq 2048, 3 epochs.

Validation per-position acceptance (pos1-7): 0.689 / 0.541 / 0.440 / 0.369 / 0.319 / 0.283 / 0.255 | full_acc 0.415.