[Feature]Use Megatron-core dist_checkpointing to load checkpoint with different parallel strategies

**Is your feature request related to a problem? Please describe.**
Currently ChatLearn use a tool to convert checkpoint when different parallel strategies are detected, https://github.com/alibaba/ChatLearn/blob/main/chatlearn/utils/megatron_utils.py#L164

The online conversion has been addressed in Megatron core dist_checkpointing. https://docs.nvidia.com/megatron-core/developer-guide/latest/api-guide/dist_checkpointing.html

**Describe the solution you'd like**
Use Megatron-core dist checkpointing to save and load checkpoint.

Start with chatlearn/utils/megatron_utils.py around line 164 and read the Megatron-core dist_checkpointing documentation linked in the issue. Trace the current checkpoint conversion path and determine how save and load should use distributed checkpointing across parallel strategies. Done means checkpoints can be saved and loaded without the existing online conversion step.

Written by the indexing model from the issue text.