Enhance recording APIs #9
Description
Right now we only capture input/output tokens specifically:
https://github.com/coder/aibridge/blob/main/api.go#L19-L25
Both OpenAI and Anthropic support prompt caching, and return this information as well, but we store it in metadata only.
We should instead be capturing the cached prompt tokens alongside input/output tokens to reflect true usage.
This is obviated by #13.
Additionally, we should be tracking the duration for both the tool invocation and the wider interception (CompletedAt to complement StartedAt).