is there any backend with smarter KV cache invalidation that llama.cpp? when I cut a few tokens at the end, it deletes the entire cache and needs to process the whole prompt from scratch