feat: add W&B worker GPU system metrics

[EazyReal]

Description

In single-controller training, the controller process initializes the W&B run while actor/rollout/critic workers own the CUDA devices. W&B samples system metrics from initialized W&B clients, so controller-only logging can show controller process memory without worker GPU utilization.

This PR adds opt-in worker-side W&B system metrics using W&B shared mode:

• stats_logger.wandb.system_metrics.enabled defaults to false, so existing runs are unchanged.
• When enabled, stats_logger.wandb.mode must be shared; the controller stays the primary run and workers attach non-primary clients.
• Worker clients are labeled by role/rank and use x_update_finish_state=False, so they contribute system metrics without finalizing the primary run.
• Worker telemetry init/finish is lock-protected, idempotent, and warning-only on worker W&B init failures.
• English/Chinese docs and generated CLI references document the opt-in settings.

Why this belongs in AReaL

This cannot be solved cleanly by a launcher-only change. The launcher can select W&B config, but AReaL owns the worker process lifecycle, role/rank identity, RPC setup, and StatsLogger run identity. Keeping this in AReaL lets the feature be role-aware, non-breaking by default, and consistent across local, Slurm, sync RPC, and Ray RPC paths.

Related Issue

None. Opening as a draft PR per CONTRIBUTING.md so maintainers can review the design before this new feature is merged. Supersedes closed setup attempts #1335, #1336, and #1337.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation update
• Code refactoring
• Performance improvement
• Test coverage improvement

Usage

stats_logger:
  wandb:
    mode: shared
    system_metrics:
      enabled: true
      roles: [actor, rollout]   # null = every configured worker role
      gpu_device_ids: null      # null = W&B samples the worker's visible GPUs

Prior PRs reviewed

Merged precedents this follows:

• feat(utils): add Trackio experiment tracking backend #1113 added a tracking backend through public config, StatsLogger lifecycle integration, tests, and docs. This PR follows that shape but adds no dependency and keeps the new path disabled by default.
• fix(api): add mode validation for WandBConfig and SwanlabConfig #1134 added early validation for tracking config modes. This PR validates system_metrics role/GPU inputs and fails early when worker metrics are enabled without wandb.mode: shared.
• docs: add checkpointing, metrics tracking, and workflow guides #891 and docs: add Important Metrics to Monitor section #894 established the metrics documentation style. This PR updates both English and Chinese metrics docs plus generated CLI references.

Closed/unmerged precedents this intentionally avoids:

• docs: fix critical doc-code mismatches across config, endpoints, LoRA, and metrics #1203 bundled unrelated documentation fixes across endpoints, rollout config, LoRA, and metrics. This PR keeps one observability feature with matching code, docs, and tests.
• fix(dataloader): prevent data drop and padding during validation for accurate metrics #1096 changed metrics semantics with an underdeveloped test/checklist story before a narrower replacement merged. This PR includes focused tests for config validation, role filtering, W&B shared settings, and cleanup behavior.
• feat(service): add chat template message normalization and IS benchmark tooling #1172 mixed a core client change with a large benchmark/tooling suite. This PR does not add benchmark tooling or unrelated workflow changes.
• Scaffolding rollout workspace multi worker #1011 was a broad multi-worker scaffolding PR. This PR only adds a narrow observability hook at existing scheduler/RPC/trainer seams.

Checklist

• I have read the Contributing Guide
• Pre-commit hooks pass (pre-commit run --all-files)
• Relevant tests pass; new tests added for new functionality
• Documentation updated and built with ./docs/build_all.sh
• Branch is up to date with main
• Self-reviewed against merged and closed upstream precedents
• This PR is a breaking change

Breaking Change Details (if applicable):

N/A. The feature is disabled by default, and existing W&B disabled, offline, and online configurations keep their current behavior.

Validation

• uv run --only-group dev python -m pre_commit run --all-files
• uv run --only-group dev python -m pytest tests/test_wandb_system_metrics.py tests/test_trackio_backend.py
• ./docs/build_all.sh (succeeds; existing docs warnings remain for algorithms/sequence_packing.md toctree inclusion and tutorial/online_proxy.md OpenClaw README xref)

Test plan

• Maintainer safe-to-test CI on the upstream 2-GPU runner.
• Optional cluster canary: run a shared-mode W&B job and confirm worker GPU utilization appears on the primary run.