Support online training without precomputing the target cache

Hi, per the [data prep docs](https://github.com/deepseek-ai/DeepSpec/blob/main/scripts/data/README.md), training requires an offline step (`prepare_target_cache.py`) that stores the target model's hidden states to disk. The docs warn this cache is _~38 TB_ for `Qwen/Qwen3-4B`, which is a blocker without large-scale storage.

Could you add an online training mode that runs the target model on the fly instead of reading from a precomputed cache? I understand it trades storage for longer training time, but it would help on storage-limited setups. Frameworks like [SpecForge](https://github.com/sgl-project/SpecForge) already support this.

Thanks!