Proposal for better cli entrypoints performance

[saltball]

Summary

I notice in entrypoints/main.py the script import all module at the beginning but actually only in main function and after args.command == statement it use few of the import.

DeePMD-kit Version

Maybe not needed

TensorFlow Version

Maybe not needed

Python Version, CUDA Version, GCC Version, LAMMPS Version, etc

No response

Details

I suggest from deepmd.entrypoints import only after args.command == statements to have maybe better cli performance for some cases.