fix some bug

[Thaurun]

Bug Fixes Summary

1. Bug 1: In the _allreduce_word_embedding_grads function, the embedding might be frozen, which can cause the code to crash.

2. Bug 2: _ParamAndGradBuffer fails to update param_start_index during initialization, potentially causing a parameter to be placed into multiple buckets. This triggers the assertion error:

   assert param not in param_gbuf_map, (
       "Param should not be in param_gbuf_map; each param only belongs "
       "to a single bucket."
   )
   

3. Bug 3: In on_save_checkpoint_success, wandb's artifact.add_reference requires absolute paths for checkpoint_path. Using relative paths will result in errors.

[xuwchen]

Thank you for providing this bug fix. We will review it as soon as possible.

[yanring]

Merged in https://github.com/NVIDIA/Megatron-LM/commit/afb755f548b48151a4408b0e9caf674b8349b589 Many thanks to the Infrastructure Center of Tencent WeChat's Technical Architecture Department(微信技术架构部-基础架构中心) for their contributions. cc @shifangx