Add W&B worker GPU system metrics

[EazyReal]

Summary

• Add stats_logger.wandb.system_metrics to opt into worker-side W&B system metrics.
• Initialize the controller W&B run as the primary shared-mode client and attach selected worker RPC processes/Ray actors as non-primary clients.
• Document the single-controller GPU telemetry setup and regenerate CLI reference docs.

Root Cause

In single-controller training the process that calls wandb.init() can run on a CPU/controller host while actor and rollout workers run on GPU nodes. W&B system metrics are collected from initialized W&B clients, so wandb.mode=online on only the controller records controller process memory but not worker GPU utilization.

Behavior

Users can enable GPU worker telemetry with:

stats_logger:
  wandb:
    mode: shared
    system_metrics:
      enabled: true
      roles: [actor, rollout]
      gpu_device_ids: null

Workers use x_primary=False, a role/rank x_label, and x_update_finish_state=False, so they contribute system metrics without writing primary run files or finishing the run.

Validation

• uv run ruff check areal/api/cli_args.py areal/utils/stats_logger.py areal/utils/wandb_system_metrics.py areal/infra/rpc/rpc_server.py areal/infra/rpc/ray_rpc_server.py areal/infra/scheduler/local.py areal/infra/scheduler/slurm.py areal/infra/scheduler/ray.py areal/trainer/rl_trainer.py areal/trainer/sft_trainer.py areal/trainer/dpo_trainer.py areal/trainer/rw_trainer.py tests/test_wandb_system_metrics.py
• uv run pytest tests/test_wandb_system_metrics.py tests/test_trackio_backend.py