refactor(dashboard): align operations workflows with PRD

[xsuler]

What does this PR do?

Refactors the operations dashboard to match the PRD information architecture and workflows across Overview, Jobs, Runtime, Task Launcher, and Agent.

Key changes include:

• Aligns page layouts, responsive sizing, light/dark themes, bilingual UI, and job-detail navigation with the PRD.
• Connects dashboard actions to backend APIs for runtime checks, repair actions, agent recovery, execution plans, and LLM-generated follow-ups.
• Improves job state handling, active-job counts, dataset paths, rollout samples, TensorBoard metrics, runtime timing, and health status.
• Adds reward, loss, grad norm, and sequence-length metric views with dense-series rendering and point tooltips.
• Adds launcher preset discovery and smoke infer/train tuning with parameters and progress reporting.
• Removes remaining dashboard mock paths where backend data is available.

see:

Related issue

Fixes #445

Type of change

• 🐛 Bug fix
• ✨ New feature
• 💥 Breaking change
• 📚 Documentation update
• ♻️ Refactoring

How was it tested?

• npm --prefix dashboard run build
• python -m pytest tests/test_dashboard_metrics_path_cpu.py -q (4 passed)
• env PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 pytest tests/test_train_cli_config_cpu.py -q (88 passed)
• Verified a flushed TensorBoard scalar is readable immediately through the dashboard metrics path.

The full CPU test suite and CUDA-dependent paths were not run locally.

Checklist

• My PR title follows the conventional commit format.
• I have linked the related issue.
• Existing tests pass (pytest tests/ -k cpu).
• New or changed behavior is covered by focused tests.
• I have described the verification performed.
• Public API changes are additive or backward compatible.

[adohe]

/lgtm