Refactor detection_tracking

[nikk-nikaznan]

• detection_utils : is it better to split the utils to train and test
• evaluate_model : especially on _load_dataset() seems redundant to the one in training ( parent class that the evaluation class and the detector class inherit from)
• refactor inference_model.py -- redundant of training end evaluate for some function
• add EXPERIMENT_NAME to args
• is there any other way to log lost function in mlflow without tensor board
• check doc_string
• Pathlib instead of os.path
• replace print statements with calls to the logger (refer evaluate.py)
• Add readme and dependencies for detection_tracking #95
• Improvements and other ideas for training detection #96
• Test for detection_tracking #91

[sfmig]

probably a good point to start : #108