fix: honor Gemini structured retry delays

[mikemikimike]

Summary

Gemini's OpenAI-compatible endpoint can return its retry cooldown in error.details[].retryDelay without a Retry-After header. The existing OpenAI retry path ignored that value and could exhaust its bounded attempts before the server-advertised cooldown.

This change:

• reads valid Google RetryInfo-style retryDelay values for 429 and 503 responses when the header is absent;
• preserves valid Retry-After precedence and existing capped, cancellation-aware waits;
• ignores malformed, negative, and non-finite durations.

Tests

• npx vitest run src/llm/provider/openai/openai-http.test.ts — 27 passed
• npx vitest run src/llm/provider/openai — 93 passed
• npm run lint
• npm run build

Validation note

The focused provider suite, type-check, and build pass locally. A full repository run on Windows completed with 6,302 of 6,317 tests passing; 11 unrelated test files failed and 4 tests were pending because of platform, tooling, and timing constraints in this environment, including ls/node ENOENT, Windows shell/PATH/path semantics, symlink expectations, and a concurrency timing case. No failure was in src/llm/provider/openai.