[codex] Account for OpenAI Responses cached tokens

[Tanisha-Katara]

Summary

• recognize OpenAI Responses-style usage fields in API cost extraction
• treat input_tokens / output_tokens as aliases for prompt / completion tokens
• read cached prompt tokens from input_tokens_details.cached_tokens as well as prompt_tokens_details.cached_tokens
• add a regression test that verifies cached-token pricing for Responses-style usage

Why

ART already accounts for OpenAI cached prompt tokens when usage is shaped like Chat Completions: prompt_tokens, completion_tokens, and prompt_tokens_details.cached_tokens.

OpenAI Responses-style usage commonly uses input_tokens, output_tokens, and input_tokens_details.cached_tokens. Without those aliases, cached input tokens can be priced as uncached input or the response can fail provider usage extraction, depending on which fields are present.

This keeps the existing pricing model and only broadens the accepted usage spellings.

Validation

• uv run pytest tests/unit/test_track_api_cost.py::TestTrackApiCost -q -> 16 passed
• uv run ruff check src/art/api_costs.py tests/unit/test_track_api_cost.py
• python3.12 -m compileall -q src/art/api_costs.py tests/unit/test_track_api_cost.py
• git diff --check

Note: the full tests/unit/test_track_api_cost.py file also runs two pipeline trainer integration tests that require torch; those failed in this fresh local uv environment because torch is not installed. The focused API-cost test class passed.

Context: this is part of a broader open-source effort to make prompt-cache costs measurable across eval and agent frameworks: https://github.com/Tanisha-Katara/cacheeconomics