DeepSeek introduces DSpark speculative decoding for 5x faster inference

DSpark, a speculative decoding system from DeepSeek, achieves up to 5x inference speedup without modifying the model. A community PR adds support to llama.cpp.