[Feature] New Speculative Decoding Framework

[Vinkle-hzt]

Due to the high CPU overhead in the existing speculative decoding framework, we are developing a brand new framework that significantly reduces CPU consumption and minimizes device-to-host synchronization.

Worklist

• [ ] support 1 & multi steps mtp [#305]
• [ ] support py model & cuda graph
• [ ] support PD-seperation
• [ ] support DP
• [ ] fast & async mtp process
• [ ] vocab prune