[Improve] Track Brain inference usage and cost

[roomote-roomote]

​Opened on behalf of @mrubens. Follow up by mentioning @roomote-roomote, in the web UI, or in Slack.

What changed

• Record every Brain provider attempt through the existing LLM usage ledger, including provider, resolved model, operation, status, and full-response latency.
• Normalize embedding, rerank, Chat Completions, and Responses token metadata from JSON and SSE responses, and distinguish their embedding, rerank, and inference usage types.
• Capture provider-reported cost as costMicroUsd with a provider_response source, while retaining failures and metadata-free responses with explicit missing-cost attribution.
• Request final usage chunks for OpenAI-compatible chat streams without replacing caller-supplied stream options.

Why this change was made

Brain inference previously passed through the provider gateway without entering Roomote's usage ledger, leaving embeddings, reranking, query expansion, and synthesis absent from cost and token reporting.

Impact

Operators can account for Brain inference alongside other LLM usage and separate it by operation, model, provider, status, and usage type. OpenRouter responses can include monetary cost; direct OpenAI responses generally provide token counts but not cost, so those rows retain costSource: missing unless the provider returns cost metadata.

[roomote-community]

No code issues found. See task

_{Reviewed 630c359}