fix: report measured OpenAI token usage - #95

Summary

Use the token usage reported by OpenAI across the ReAct loop instead of estimating usage from the number of messages.

The provider currently reports:

total: result.messages.length * 100
prompt: Math.floor(result.messages.length * 60)
completion: Math.floor(result.messages.length * 40)

and derives cost from a hard-coded per-token rate. Two runs with the same message count therefore report identical usage even when the API reports different token counts, and the cost cannot be correct for arbitrary configured models or OpenAI-compatible base URLs.

Changes

• aggregate completion.usage across every successful ReAct completion
• include the number of OpenAI requests in the aggregated usage
• return measured usage through the Promptfoo provider response
• stop emitting synthetic cost when the provider has no authoritative price source
• add regressions for a two-request tool loop and provider response passthrough

If any successful completion omits usage, the aggregate is omitted rather than presenting a partial total as complete.