feat: recognize provider cost data from usage.cost_details block

[mcowger]

Summary

Some providers include detailed cost breakdowns directly in the response usage object rather than via SSE : cost comment lines. This PR adds support for recognizing and using that cost data when available.

The new format

Providers may return a usage block like:

"usage": {
  "prompt_tokens": 23,
  "total_tokens": 66,
  "completion_tokens": 43,
  "cost": 0.00017465,
  "cost_details": {
    "total_cost": 0.00017465,
    "input_cost": 0.00002415,
    "output_cost": 0.0001505,
    "cached_input_cost": 0,
    "cache_write_input_cost": 0,
    "upstream_inference_cost": 0.00017465,
    "request_cost": 0,
    "web_search_cost": 0,
    "data_storage_cost": 0.00000106
  }
}

When cost_details is present, we use the provider's actual per-bucket breakdown (input_cost, output_cost, cached_input_cost, cache_write_input_cost) directly instead of proportionally distributing a total.

Changes

• utils/usage-normalizer.ts: Add extractUsageCostDetails() — safely extracts cost_details from provider usage blocks; returns null when absent (providers that don't use this format are unaffected). Also updated normalizeOpenAIChatUsage() to extract cache_write_tokens from prompt_tokens_details.
• utils/provider-cost.ts: Add applyUsageCostDetails() — applies per-bucket breakdown when available, falls back to proportional distribution otherwise.
• services/inspectors/usage-logging.ts: Wire cost_details extraction into the streaming cost path (only applies if no SSE-reported cost was found).
• services/response-handler.ts: Same for the non-streaming (unary) path.

Key design decisions

• Fully optional/defensive: extractUsageCostDetails() returns null when usage.cost_details doesn't exist — providers that don't use this format are completely unaffected
• SSE : cost comments take precedence: The !providerReportedCost guard ensures cost_details only applies when no SSE-reported cost was found
• Per-bucket breakdown preferred: When the provider gives explicit input_cost/output_cost/etc., we use those directly instead of proportional splitting

Test plan

• extractUsageCostDetails — extracts from the new format, falls back to usage.cost/usage.estimated_cost, returns null for missing/invalid data
• applyUsageCostDetails — uses per-bucket breakdown, falls back to proportional, handles zero/null costs
• normalizeOpenAIChatUsage — extracts cache_write_tokens from prompt_tokens_details
• Precedence: SSE : cost comments > cost_details > calculated costs
• All 1367 existing tests pass