add merge_lora.py

[Nsigma-Bill]

This PR:

1. Add a new script named merge_lora.py
2. Update and simplify the previous LoRA merging process (perform merging operation under the new huggingface peft framework and remove unnecessary operations in export_*_checkpoint.py).
3. Make the codes more straightforward and easier to read.
4. The resulting merged standalone model bin files could be directly used for llama.cpp conversion.