In this Pull Request, in order to facilitate the dynamic update of pytorch_config.yaml (when calling deeplabcut.train_network()), I:

• Implement the update_config_by_dotpath() method. This method is easier to use than the previous update_config(). With update_config(), you needed to pass a nested dictionary as argument to target nested arguments in the pytorch_config.yaml, and it wasn't explicit how the nested dict was processed. With update_config_by_dotpath(), you target nested arguments by using a single-level dictionary with dot-notation path access. For example to update the gpus list in the runner section, you pass {"runner.gpus": [0,1]} (instead of dict(runner=dict(gpus=[0,1])))
• Pass updates to the pytorch configuration with the pytorch_cfg_updates: dict argument instead of using the kwargs in train_network().
• Make explicit all the other arguments that can be passed to train_network() (that were passed through the kwargs before) for more transparency, and remove the kwargs argument from train_network().
• Implement a test for the new update_config_by_dotpath() method, update integration tests.