updating docs to use fork of megatron-lm and minor example/docs fix

[pacman100]

What does this PR do?

1. Maintaining 🤗 fork of Megatron-LM: https://github.com/huggingface/Megatron-LM. Updating docs to directly install from it instead of the current long process.
2. wandb was resulting in the below error when doing accelerator.end_training() with Megatron-LM. On commenting those 2 lines, everything works properly. For Megatron-LM, the main process is the last rank as that will have loss from the last pipeline stage, this might be behaving weirdly with closing the wandb connection using run.finish. Commenting it for now.

envs/ml/lib/python3.10/site-packages/wandb/sdk/service/streams.py", line 187, in get_stream
    stream = self._streams[stream_id]
KeyError: '26ll3a5t'

3. Minor megatron-lm docs fixes to make output logs in sync with the config and args

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[sgugger]

Thanks for the fixes! Do you plan to update the fork regularly?

[pacman100]

Hello @sgugger, wrt maintaining fork, sure, I can do that. This would help in making sure the functionality works with accelerate and we can extend it for more features (main repo is barely active).