Log rate-limit headers and cache usage

The script is small: 20 calls with the same cached system prompt, each through messages.with_raw_response.create so the headers and the parsed usage arrive together, with max_retries=0 so one call is one HTTP request. The decision that matters most is where you point it. OpenRouter and Ollama speak the Messages API but don't pass on Anthropic's per-organisation anthropic-ratelimit-* headers, so the project includes a local fake provider that meters ITPM the way the Claude API does and reports it in those headers. On the fake you see cache-aware ITPM in the headers. On OpenRouter with an Anthropic model you see the same effect in the usage fields, at a cost of well under a cent.