[Feature Request] MOBA：Mixture of Block Attention For Long-context

**Summary**
Request implementation of MoBA Attention, the technique used by Kimi for efficient long-context training and inference.

**Current state**
Kimi official has already placed the basic code in their repository. Are there any plans to integrate it into megatron-lm？
MoBA：https://github.com/MoonshotAI/MoBA
