[Perf][MTP] Improve speculative decoding efficiency

[freeliuzc]

Reduced kernel launch latency:

• The original custom op library was too large, causing kernel launch latency to reach ~200µs. By using PyBind instead, launch latency was reduced to ~20µs.

• Synchronization improvements:
  Optimized several sync points in the speculative decoding path.

Conclusion：
In the ERNIE-LITE + MTP pipeline, the time from verify to the next main model C-Embedding was reduced from 8.5ms to 4ms(113%).

[paddle-bot]

Thanks for your contribution!