Create a small tutorial on how to accelerate HF Llama models with Transformer-Engine

[sudhakarsingh27]

1. Add a te_llama.py file which has following code:
   • TELlamaDecoderLayer (wrapper over TE's TransformerLayer which replace HF's LlamaDecoderLayer)
   • TELlamaForCausalLM creates the Language Model with TELlamaDecoderLayer instead of LlamaDecoderLayer.
   • from_pretrained_local that loads HF Llama 2 checkpoint weights (which are meant for LlamaDecoderLayer) into TELlamaDecoderLayer (ultimately the TE's TransformerLayer).
2. Add a tutorial in jupyter notebook tutorial_accelerate_hf_llama_with_te.ipynb that showcases how to use the new TELlamaDecoderLayer. Also shows some basic benchmarks on H100 GPUs.
3. Add utils.py file which contains necessary dataloading, model setup code to run the tutorial notebook seamlessly.

[ptrendx]

Please add a link to the tutorial to https://github.com/NVIDIA/TransformerEngine/blob/main/docs/index.rst

[timmoon10]

😎

[sudhakarsingh27]

😎

[ptrendx]

From my side LGTM, thank you!