DeepSeek-V4-Flash-DSpark

DeepSeek released [DeepSeek-V4-Flash-DSpark variant](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-DSpark) of the V4 Flash model with built-in speculative decoding and described the methodology in a [companion paper](https://github.com/deepseek-ai/DeepSpec/blob/main/DSpark_paper.pdf). Given the focus on performance of running the model locally, I think it makes sense to explore how DSpark can be integrated into ds4.

It seems that DSpark provides considerable improvement in tok/s performance with fixed concurrency over previous generation of speculative decoding (MTP):

Start with the linked DeepSeek-V4-Flash-DSpark model page and DSpark paper to understand the proposed speculative-decoding method. Then inspect ds4's C implementation and current inference paths to identify where DSpark could fit; done means a working integration with tok/s performance compared against the existing MTP approach.

Scritto dal modello di indicizzazione a partire dal testo della issue.