Support online training without precomputing the target cache

Hi, per the data prep docs, 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 already support this.

Thanks!