potential non-negligible slowdowns from the introduction of graph breaks while using cache-dit with torch.compile. Also see: https://github.com/huggingface/flux-fast/pull/13 https://github.com/vipshop/cache-dit/blob/b954809d32a152e9077c888862ea6890a1a9f2bb/src/cache_dit/cache_factory/dual_block_cache/cache_context.py#L1163