fix: auto-log Tinker inference costs from ART managed openai_client

[vivekkalyan]

Summary

• auto-log costs/{train,eval}/tinker_prefill and costs/{train,eval}/tinker_sample for calls made through ART's managed model.openai_client()
• enable the same pricing-backed cost calculator setup in LocalBackend.register() when model pricing exists
• add focused unit coverage for train, eval, no-context, and multi-choice inference cost logging

Why

Internal repo's PipelineTrainer path was using ART-managed inference but was not emitting Tinker rollout costs unless the codebase manually called model.cost_calculator(...). This moves that responsibility into ART so repos do not need rollout-side plumbing.

Notes

• this keeps the cleaner contract: ART-managed openai_client() calls auto-log costs when an ART metrics context is active
• downstream repos that already add the same Tinker costs manually should remove that duplicate logging

Other option

• Introduce a first-class ART inference API like model.generate(...) or model.chat_completions(...).
• That API owns cost/timing/token accounting.
• openai_client() stays raw.
• Cleanest, but requires codebase migration.

Validation

• .venv/bin/python -m pytest tests/unit/test_model_openai_client_costs.py tests/unit/test_track_api_cost.py -q