fix(usage): use OpenRouter's billed cost instead of estimating locally

[srtab]

For OpenRouter (DAIV's default Anthropic route), langchain_openai's usage parser only emits cache_read, never cache_creation. The previous code passed cache_write_tokens=0 to genai_prices, which then priced cache-creation tokens at the base input rate instead of the 1.25x/2x cache-write rate. With the default 1h cache TTL, cache writes were being undercharged by about 50%, producing run costs visibly lower than LangSmith's.

Switch to OpenRouter's authoritative cost field:

• ChatOpenRouter subclasses ChatOpenAI and stashes usage.cost on the message's response_metadata so it survives streaming chunk merging (non-streaming already exposes it via llm_output propagation).
• get_model routes OpenRouter calls through ChatOpenRouter and sets extra_body.usage.include + stream_usage so OpenRouter actually returns the cost field.
• DaivUsageCallbackHandler aggregates per-model provider cost alongside the existing token usage; build_usage_summary prefers it over the local genai_prices estimate and tags each model entry with cost_source.

Direct Anthropic / OpenAI / Google keep using genai_prices unchanged.