Reslove multi-gpu training error and add description of dataset format and label format

[shyhyawJou]

if not remove the decorator rank_zero_only, it will encounter an error as shown below, when the users use multi-gpu training and turn-on tensorboard.

Traceback (most recent call last):
  File "/media/user/disk2/mateo/ramen/YOLO-mainz/yolo/lazy.py", line 17, in main
    callbacks, loggers, save_path = setup(cfg)
  File "/media/user/disk2/mateo/ramen/YOLO-mainz/yolo/utils/logging_utils.py", line 286, in setup
    loggers.append(TensorBoardLogger(log_graph="all", save_dir=save_path))
  File "/media/user/disk2/mateo/py310_venv/sushilon/lib/python3.10/site-packages/lightning/pytorch/loggers/tensorboard.py", line 96, in __init__
    super().__init__(
  File "/media/user/disk2/mateo/py310_venv/sushilon/lib/python3.10/site-packages/lightning/fabric/loggers/tensorboard.py", line 98, in __init__
    root_dir = os.fspath(root_dir)
TypeError: expected str, bytes or os.PathLike object, not NoneType

[sam31046]

I have this issue whatever I set device={cpu/0/[0,1]}. I believe it's because the save_path is not shared between multi-process. save_path is not defined in non-main process. They instead skip the validate_log_directory line in setup function.

I think this PR is a temporary fix. We need to refactor setup a little bit. I'm not familiar with broadcast though.
Check out this DDP discussion, Lightning-AI/pytorch-lightning/issues/18148 @henrytsui000

Helpful way to see who's calling validate_log_directory:

@rank_zero_only 
def validate_log_directory(cfg: Config, base_path: str, exp_name: str): 
    print(f"PID={os.getpid()}, LOCAL_RANK={os.environ.get('LOCAL_RANK', 'NA')}, RANK={os.environ.get('RANK', 'NA')}") 
    print(f"[DEBUG] Called validate_log_directory from rank_zero_only, return will be valid only on rank 0.")