Replace text_encoder with google/gemma-3-12b-it (bf16, 24GB vs 49GB float32)

Convert text_encoder weights from float32 to bfloat16 (49GB → 27GB)

Add model.safetensors.index.json for transformers compatibility