Metric Logging - Enable kwargs so user can add extra args

[felipemello1]

Before we would pass a whole dictionary as input to a logger backend and pop a selected few args

class LoggerBackend(ABC):
	def __init__(self, logger_backend_config: dict[str, Any]) -> None:
		self.project = logger_backend_config["project"]

This means that if users want to pass extra args, e.g. run name or tags, they couldnt.

Now we use **kwargs and pass it directly to the wandb run:

class LoggerBackend(ABC):
	def __init__(self, *, logging_mode, per_rank_share_run, **kwargs):
           self.backend_kwargs = kwargs

	...

	async def _init_per_rank(self, run_name: str):
        self.run = wandb.init(name=run_name, **self.backend_kwargs)

Tested it with multiple args. We can even pass a custom name. e.g.

logging_mode: global_reduce

logging_mode: per_rank_reduce

[wukaixingxp]

LTGM, tested this branch with my yaml:

metric_logging:
  wandb:
    entity: "torchforge"
    project: "kaiwu-coding-grpo"
    logging_mode: global_reduce
  console:
    logging_mode: global_reduce

and saw
wandb: 🚀 View run royal-feather-3 at: https://meta.wandb.io/torchforge/kaiwu-coding-grpo/runs/8xpli7ro wandb: Find logs at: wandb/run-20251021_153007-8xpli7ro/logs

[wukaixingxp]

LGTM, thanks for helping me on the wandb issue!