fix: reduce max_tokens and add retries to handle Groq TPM rate limits gracefully

fix: switch default model to llama-3.1-8b-instant to bypass 70b daily quota limit

fix: reduce stop sequences to 4 to comply with Groq API limits

fix: bind stop sequences to llm to prevent hallucination loop and rate limit 429 errors

feat: implement Madia multi-agent debate app

chore: add project structure with .gitignore