Add torchtune convertor and README changes

[kartikayk]

torchtune supports HF and Meta formatted checkpoints. In this PR I'm add a custom convertor for torchtune. Currently this only support llama2 7B and llama3 8B. I also update the README with the necessary steps.

Test:

Llama3 Model:

python3 build/convert_torchtune_checkpoint.py \
--checkpoint-dir /data/users/kartikayk/cpts/Meta-Llama-3-8B/ \
--checkpoint-files meta_model_0.pt \
--model-name llama3_8B \
--checkpoint-format meta

python3 torchchat.py generate \
--checkpoint-path /data/users/kartikayk/cpts/Meta-Llama-3-8B/model.pth \
--device cuda

Output:

HF Format Llama2 7B

python3 build/convert_torchtune_checkpoint.py \
--checkpoint-dir /tmp/Llama-2-7b-hf/ \
--checkpoint-files hf_model_0001_0.pt hf_model_0002_0.pt \
--model-name llama2_7B \
--checkpoint-format hf

python3 torchchat.py generate --checkpoint-path /tmp/Llama-2-7b-hf/model.pth  --device 'cuda'

Output:

[mikekgfb]

Thank you!