Remove HF_TOKEN dependency in E2E test

[jack8558]

Removing HF_TOKEN dependency in E2E test

• Created tp save_hf_model_files_to_gcs to save huggingface model files in gcs
• Saved tokenizers and llama3-8B's model weights and configs in gcs bucket (the weights and configs are needed for SFT e2e test)
• Since huggingface library can't load directly from gcs, added util funciton copy_gcs_to_local which download in tmp directory
• Removed HF_TOKEN on e2e test and cpu test

tp save_hf_model_files_to_gcs example:

tp save-hf-model-files-to-gcs \
  --repo-id "meta-llama/Meta-Llama-3-8B" \
  --gcs-path "gs://bucket" \
  --file-type "all" \
  --temp-dir /mnt/disks/tmp

#14