fix(text_encoder): remove orphaned index, update generation_config

fix(text_encoder): restore Lightricks config with layer_types and full architecture fields

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

Upload folder using huggingface_hub