Fix Japanese Hugging face GPT conversion

[noppayut]

Fixes #459

1. Define missing variables, correct variable names.
2. Use GPT2LMHeadModel and state_dict() instead of GPT2Model and named_parameters() to collect model weights.
3. Dynamically select torch device instead of the hardocded GPU.
4. Add tensor_para_size to config.
5. Add docstring to recommend usage of huggingface_gpt_model.py instead of this file.