[MPK] Support Speculative Decoding

[Risc-lt]

Description

This issue tracks MPK's explorations for supporting speculative decoding

Multi-head speculation

1. Single draft sequence
   1. Support EAGLE3 for Qwen model series [MPK] Support EAGLE3 Speculative Decoding for Qwen3 #692
   2. Refactor Multi-head pipeline for generality and performance
      1. Separate draft kvcache pool with unified page memory allocator
      2. Overlap draft generation and verification
      3. Reorganize KNGraph with additional dependency, e.g. h_norm / e_norm
2. Multiple draft candidates
   1. Tree attention kernel
   2. Attention planning
   3. Dynamic verification budgets