Reduce training to 500 steps with tightened curriculum for A10G budget

Fix eval device selection with CUDA-safe fallback

Optimize for A100 80GB: 8 generations, batch 4, lr 2e-5, dense logging

Restore full 1000-step training with original curriculum

Reduce training to 500 steps with tightened curriculum for A10G budget

Add Gradio training monitor and fix subprocess python path

Update: Started making changes for the hackathon