Fix train_ssd.py: use correct model class (AutoModelForImageTextToText), target language_model only for LoRA, fix deprecated torch_dtype, remove OOM-causing prepare_model_for_kbit_training

Add SSD training script for Gemma-4-E2B-it