fix(agent): handle non-dict LLM response gracefully and delay setting task to FAILED during Celery retries

revert LLMService back to official Google GenAI SDK

refactor LLMService to route Gemini requests through OpenRouter

feat: add prometheus metric tracking for token counts, agent latency, and failure rates

chore: upgrade default gemini model to 2.5-flash and remove hardcoded models

feat: implement google gemini llm helper service with structured json configuration

feat: added files and folders