Bug of MoE checkpoint.

Now the parallelism of attention and moe is decouple. So there are so many bugs in convert megatron checkpoint to hf checkpoint, especially the expert_tensor_parallel. Please arrange for someone to handle it.

Start by locating the Megatron-to-HF checkpoint conversion entry point and tracing how expert_tensor_parallel is handled when attention and MoE parallelism are decoupled. Done means the conversion no longer exhibits the reported MoE checkpoint bugs and produces a usable HF checkpoint; the issue does not name specific files or tests.

Written by the indexing model from the issue text.