Add Modal GPU inference support for faster LLM responses

Phase 1: project setup, MiniMax-M3 synthetic dataset generation, 1446-row clean dataset