feat(tracking): plumb wandb tags through Tracking init

[dinhxuanvu]

Why

Tracking doesn't currently pass tags to wandb.init, so SkyRL-driven runs are always untagged. Tags are W&B's primary primitive for filtering, grouping, and organizing runs in the UI; not having them means callers can't distinguish, e.g., "baseline vs. ablation," "SFT vs. RL," or sibling runs in a sweep.

This also closes a small surprise gap: the lower-level WandbTracker in skyrl/utils/log.py already plumbs arbitrary kwargs to wandb.init, so power users discovered they could set tags through that path but not through the supported Tracking adapter.

What

• Tracking.__init__ accepts a new explicit tags: Optional[List[str]] = None parameter.
• The wandb branch forwards it to wandb.init(tags=tags).
• New trainer.tags config field (default None).
• get_tracker in main_base.py forwards cfg.trainer.tags.
• The SFT path mirrors this: SFTConfig.tags bridges to cfg.trainer.tags, and SFTTrainer._init_tracker forwards it.
• Tags are wandb-specific; other backends are unchanged.

Why an explicit param

Tracking is an adapter over five backends (wandb, mlflow, swanlab, tensorboard, console). A **kwargs passthrough would couple the adapter to wandb's API surface. An explicit, typed parameter keeps the contract clear and leaves a clean place to add matching primitives (mlflow tags, etc.) without breaking changes.

Test plan

• Unit test: Tracking(..., tags=["a","b"]) results in wandb.init(..., tags=["a","b"]).
• Unit test: default tags=None is forwarded when not passed.
• Format/lint clean.
• No behavior change when tags is not passed.

Compatibility

Strictly additive. wandb.init(tags=None) is a no-op; the SDK still honors WANDB_TAGS env var.

[gemini-code-assist]

Code Review

This pull request introduces Mixture of Experts (MoE) auxiliary-loss metrics collection and logging, a background GPU/RAM monitor (RayGpuMonitor) that scrapes Ray node Prometheus endpoints, and improved exception logging to wandb. Feedback focuses on correcting the type of moe_aux_loss_coeff to float to prevent truncation, safeguarding against potential ZeroDivisionError when self._num_training_gpus is zero, and using a threading.Event instead of a boolean flag and time.sleep to allow the background monitoring thread to shut down immediately.

[dinhxuanvu]

Please ignore the Gemini's review as there were some unrelated commits which were pushed accidentally and Gemini reviewed polluted code which was incorrect. These were removed.

[SumanthRH]

Can you resolve merge conflicts @dinhxuanvu ?