fix: strictly bound task scores to satisfy OpenEnv Phase 2 testing bounds

feat: switch default API to Groq (llama-3.3-70b-versatile) for faster inference

fix: use Qwen2.5-Coder-32B-Instruct (available on HF router) instead of Qwen3-1.7B

fix: rewrite inference.py & client.py to match OpenEnv sample pattern

feat: add inference.py, openenv.yaml, bug dataset & openai dep for hackathon submission