Bridging Draft Policy Misalignment: Group Tree Optimization for Speculative Decoding