[fix][train] Defer W&B authentication to SDK

[bvolpato]

Summary

SkyRL checked only WANDB_API_KEY before calling W&B. That rejected credentials stored by wandb login and offline runs that do not need API access. It also replaced W&B's authentication errors with a less useful SkyRL assertion.

The tracker is created inside a Ray task, so the W&B SDK controls used by the launcher must also reach that task.

Changes

• Remove the environment-only check from training configuration validation.
• Remove the same check from WandbTracker.
• Forward WANDB_API_KEY, WANDB_MODE, and WANDB_BASE_URL to the Ray runtime when they are set.
• Let wandb.init select and validate the configured authentication mode.

API-key, netrc, offline, and custom-endpoint authentication now follow W&B SDK behavior.

Validation

• Config, tracking, logging, and runtime-environment tests (186 passed)
• Regression coverage for W&B settings that are set and unset in the Ray runtime environment
• W&B SDK and Tracking initialization in offline mode without WANDB_API_KEY
• Training run authenticated through netrc and synchronized metrics successfully
• uv lock --check
• Scoped Ruff, Black, and hardcoded-secret checks
• git diff --check

[gemini-code-assist]

Code Review

This pull request removes explicit checks for the WANDB_API_KEY environment variable in validate_generator_cfg and WandbTracker, deferring authentication directly to the wandb library. Unit tests have been added to ensure that the configuration validation and tracker initialization succeed even when the environment variable is not set. I have no feedback to provide as the changes are clean and well-tested.