MagicDec-part2 MagicDec: Breaking the Latency-Throughput Tradeoff for Long Contexts with Speculative Decoding