Change llama 3 and 2 configs to use hf ckpt instead of meta ckpt

[felipemello1]

Update llama 3 and 2 to use hf ckpt, so users can use the trained model using .from_pretrained.

3.1 reference:

torchtune/recipes/configs/llama3_1/8B_full.yaml

Line 39 in e9fd56a

_component_: torchtune.training.FullModelHFCheckpointer

3.0 currently:

torchtune/recipes/configs/llama3/8B_full.yaml

Line 39 in e9fd56a

_component_: torchtune.training.FullModelMetaCheckpointer

steps:

1. Change the checkpointer FullModelMetaCheckpointer -> FullModelHFCheckpointer
2. Update the download command to --ignore-patterns "original/consolidated.00.pth", instead of safetensors
3. Update checkpoint files to .safetensors
4. Launch training without errors