feat: Improved metrics in ART

[vivekkalyan]

Summary

This PR implements the W&B metrics taxonomy RFC in ART so run metrics land in stable top-level namespaces that are easier to panel and compare in W&B.

The main behavior change is hierarchical cost logging. Users can log leaf metrics like costs/train/llm_judge/correctness or costs/train/tinker_train, and ART will automatically emit parent rollups like costs/train and costs/all, plus cumulative costs/cum/* metrics across steps.

This PR also adds a MetricsBuilder API and a @track_api_cost(...) decorator so judge / external API spend can be logged into the same taxonomy with model-aware pricing.

What changed

• Canonicalized ART/W&B metric sections to reward/*, loss/*, throughput/*, costs/*, time/*, data/*, train/*, val/*, and test/*.
• Routed backend-emitted metrics into those namespaces across local, serverless, unsloth, and tinker paths.
• Added MetricsBuilder for user-owned metrics, hierarchical cost rollups, costs/cum/*, cumulative time/* and data/*, derived throughput metrics, exact data/cum/num_unique_scenarios, and persisted resume state.
• Added automatic ART-owned metrics where available, including time/step_wall_s, time/step_trainer_s, time/step_actor_s, time/step_eval_s, data/step_*, train/*, and automatic local costs/gpu when pricing is known.
• Added @track_api_cost(...) for OpenAI/Anthropic responses with explicit provider and model_name, cache-aware pricing, and hooks for custom extractors / registered pricing.
• Updated W&B metric definitions and docs for the new taxonomy.
• Removed support for legacy costs_* keys in favor of hierarchical costs/... paths.

Notes

• track_api_cost(...) requires explicit provider and model_name and raises if pricing is missing
• ART still uses training_step as the x-axis;
• But, we can now change the x-axis to metrics other than training_step (e.g. costs)