DeepSeek open-sources inference optimizations with 60-85% speedup

DeepSeek's DSpark paper describes inference optimizations achieving 60-85% faster generation. The method is open-sourced under the deepseek-ai/DeepSpec repository.