Fix LoRA tensor transpose + HF Gemma 3 path mapping

Rewrite adapter_config.json in PEFT format (peft_type=LORA, r=8, alpha=160)

Add PEFT-format adapter (mlx-lm naming converted to base_model.model.*.lora_A/B)

Link to HF Space playground

v1 LoRA: 89% accuracy on 150-Q 11+ holdout, -43 Elo vs Gemma 3 4B base

initial commit