Refine setup process

[Yuqing]

• manage default devices configuration
  • change name to predictors.yaml
  • include configs/predictors.yaml in setup.py, refer to https://setuptools.readthedocs.io/en/latest/userguide/datafiles.html
  • copy to user home directory (~/.nn-meter/configs/predictors.yaml)
  • get_config when testing and using
• clarify / check dependencies
  • my suggestion is to exclude frameworks (e.g. tensorflow / pytorch / nni) from installation dependencies (because user may only use one of them)
  • check the framework installation and version on demand when user imports models
• add entry_point to support command line usage
• downloaded predictors (data / test data) to ~/.nn_meter/data
• add --verbose to print debug-level information (e.g., the divided kernel information)
  • refine logging level
  • use two buffer to hold logging
• add --tersorflow etc to specify model type
• try model compression with gzip
• Specify parameter type in the function declaration
• change nn_meter/utils/graphe_tool.py to graph_tool.py and Graph class
  • change all graphe to graph and check its using.
• add requirements in setup.py
• code testing
  • .pb model
  • .onnx model
  • link an action on GitHub with .yml
  • setting cache policy to avoid redundancy downloading in test (Leaving to the next PR)