Refine benchmark cost and time accounting

[Qian-Cheng-nju]

Summary

• record formal agent wall time and equivalent public API cost for every benchmark backend
• use native agent telemetry where available and maintained public pricing otherwise
• keep infra and quota attempts in separate diagnostic accounting instead of formal benchmark totals
• report missing or incomplete cost as unavailable rather than zero

[munimthahmid]

@Qian-Cheng-nju Reviewed the implementation thoroughly. Also I ran Codex, Claude Code, and one shot OpenAI calls on a small proof, including complete and timeout paths . Wall time, token/request accounting, fail-closed behavior, and headline equivalent costs were reported correctly.

I found no major issue. Approving.

[Qian-Cheng-nju]

Thanks a lot! @munimthahmid Merging.