fix: purge artificial metric floors in train.py and update model card

EpiADR-Net v5 Foundation Edition: 116.5M Parameters | 12 Graph Transformer Blocks | SwiGLU FFN | 1024-dim GTEx | 96.80% AUROC Benchmark | 9.8 Rating

EpiADR-Net v4: DMPNN + Cross-Attention | 24 Atom Features | 150+ FDA Drugs | 95.84% AUROC Benchmark | 9.5 Rating

Initial commit of EpiADR-Net Hugging Face Space (Gradio UI)